1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 /* Whether the swap controller is active */
76 #ifdef CONFIG_MEMCG_SWAP
77 int do_swap_account __read_mostly;
79 #define do_swap_account 0
82 static const char * const mem_cgroup_stat_names[] = {
91 static const char * const mem_cgroup_events_names[] = {
98 static const char * const mem_cgroup_lru_names[] = {
107 * Per memcg event counter is incremented at every pagein/pageout. With THP,
108 * it will be incremated by the number of pages. This counter is used for
109 * for trigger some periodic events. This is straightforward and better
110 * than using jiffies etc. to handle periodic memcg event.
112 enum mem_cgroup_events_target {
113 MEM_CGROUP_TARGET_THRESH,
114 MEM_CGROUP_TARGET_SOFTLIMIT,
115 MEM_CGROUP_TARGET_NUMAINFO,
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
122 struct mem_cgroup_stat_cpu {
123 long count[MEM_CGROUP_STAT_NSTATS];
124 unsigned long events[MEMCG_NR_EVENTS];
125 unsigned long nr_page_events;
126 unsigned long targets[MEM_CGROUP_NTARGETS];
129 struct reclaim_iter {
130 struct mem_cgroup *position;
131 /* scan generation, increased every round-trip */
132 unsigned int generation;
136 * per-zone information in memory controller.
138 struct mem_cgroup_per_zone {
139 struct lruvec lruvec;
140 unsigned long lru_size[NR_LRU_LISTS];
142 struct reclaim_iter iter[DEF_PRIORITY + 1];
144 struct rb_node tree_node; /* RB tree node */
145 unsigned long usage_in_excess;/* Set to the value by which */
146 /* the soft limit is exceeded*/
148 struct mem_cgroup *memcg; /* Back pointer, we cannot */
149 /* use container_of */
152 struct mem_cgroup_per_node {
153 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
157 * Cgroups above their limits are maintained in a RB-Tree, independent of
158 * their hierarchy representation
161 struct mem_cgroup_tree_per_zone {
162 struct rb_root rb_root;
166 struct mem_cgroup_tree_per_node {
167 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
170 struct mem_cgroup_tree {
171 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
174 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
176 struct mem_cgroup_threshold {
177 struct eventfd_ctx *eventfd;
178 unsigned long threshold;
182 struct mem_cgroup_threshold_ary {
183 /* An array index points to threshold just below or equal to usage. */
184 int current_threshold;
185 /* Size of entries[] */
187 /* Array of thresholds */
188 struct mem_cgroup_threshold entries[0];
191 struct mem_cgroup_thresholds {
192 /* Primary thresholds array */
193 struct mem_cgroup_threshold_ary *primary;
195 * Spare threshold array.
196 * This is needed to make mem_cgroup_unregister_event() "never fail".
197 * It must be able to store at least primary->size - 1 entries.
199 struct mem_cgroup_threshold_ary *spare;
203 struct mem_cgroup_eventfd_list {
204 struct list_head list;
205 struct eventfd_ctx *eventfd;
209 * cgroup_event represents events which userspace want to receive.
211 struct mem_cgroup_event {
213 * memcg which the event belongs to.
215 struct mem_cgroup *memcg;
217 * eventfd to signal userspace about the event.
219 struct eventfd_ctx *eventfd;
221 * Each of these stored in a list by the cgroup.
223 struct list_head list;
225 * register_event() callback will be used to add new userspace
226 * waiter for changes related to this event. Use eventfd_signal()
227 * on eventfd to send notification to userspace.
229 int (*register_event)(struct mem_cgroup *memcg,
230 struct eventfd_ctx *eventfd, const char *args);
232 * unregister_event() callback will be called when userspace closes
233 * the eventfd or on cgroup removing. This callback must be set,
234 * if you want provide notification functionality.
236 void (*unregister_event)(struct mem_cgroup *memcg,
237 struct eventfd_ctx *eventfd);
239 * All fields below needed to unregister event when
240 * userspace closes eventfd.
243 wait_queue_head_t *wqh;
245 struct work_struct remove;
248 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
249 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
252 * The memory controller data structure. The memory controller controls both
253 * page cache and RSS per cgroup. We would eventually like to provide
254 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
255 * to help the administrator determine what knobs to tune.
257 * TODO: Add a water mark for the memory controller. Reclaim will begin when
258 * we hit the water mark. May be even add a low water mark, such that
259 * no reclaim occurs from a cgroup at it's low water mark, this is
260 * a feature that will be implemented much later in the future.
263 struct cgroup_subsys_state css;
265 /* Accounted resources */
266 struct page_counter memory;
267 struct page_counter memsw;
268 struct page_counter kmem;
270 /* Normal memory consumption range */
274 unsigned long soft_limit;
276 /* vmpressure notifications */
277 struct vmpressure vmpressure;
279 /* css_online() has been completed */
283 * Should the accounting and control be hierarchical, per subtree?
289 atomic_t oom_wakeups;
292 /* OOM-Killer disable */
293 int oom_kill_disable;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds;
304 /* For oom notifier event fd */
305 struct list_head oom_notify;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock;
318 struct task_struct *move_lock_task;
319 unsigned long move_lock_flags;
323 struct mem_cgroup_stat_cpu __percpu *stat;
325 * used when a cpu is offlined or other synchronizations
326 * See mem_cgroup_read_stat().
328 struct mem_cgroup_stat_cpu nocpu_base;
329 spinlock_t pcp_counter_lock;
331 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
332 struct cg_proto tcp_mem;
334 #if defined(CONFIG_MEMCG_KMEM)
335 /* Index in the kmem_cache->memcg_params.memcg_caches array */
337 bool kmem_acct_activated;
338 bool kmem_acct_active;
341 int last_scanned_node;
343 nodemask_t scan_nodes;
344 atomic_t numainfo_events;
345 atomic_t numainfo_updating;
348 /* List of events which userspace want to receive */
349 struct list_head event_list;
350 spinlock_t event_list_lock;
352 struct mem_cgroup_per_node *nodeinfo[0];
353 /* WARNING: nodeinfo must be the last member here */
356 #ifdef CONFIG_MEMCG_KMEM
357 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
359 return memcg->kmem_acct_active;
363 /* Stuffs for move charges at task migration. */
365 * Types of charges to be moved.
367 #define MOVE_ANON 0x1U
368 #define MOVE_FILE 0x2U
369 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
371 /* "mc" and its members are protected by cgroup_mutex */
372 static struct move_charge_struct {
373 spinlock_t lock; /* for from, to */
374 struct mem_cgroup *from;
375 struct mem_cgroup *to;
377 unsigned long precharge;
378 unsigned long moved_charge;
379 unsigned long moved_swap;
380 struct task_struct *moving_task; /* a task moving charges */
381 wait_queue_head_t waitq; /* a waitq for other context */
383 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
384 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
388 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
389 * limit reclaim to prevent infinite loops, if they ever occur.
391 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
392 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
395 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
396 MEM_CGROUP_CHARGE_TYPE_ANON,
397 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
398 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
402 /* for encoding cft->private value on file */
410 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
411 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
412 #define MEMFILE_ATTR(val) ((val) & 0xffff)
413 /* Used for OOM nofiier */
414 #define OOM_CONTROL (0)
417 * The memcg_create_mutex will be held whenever a new cgroup is created.
418 * As a consequence, any change that needs to protect against new child cgroups
419 * appearing has to hold it as well.
421 static DEFINE_MUTEX(memcg_create_mutex);
423 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
425 return s ? container_of(s, struct mem_cgroup, css) : NULL;
428 /* Some nice accessors for the vmpressure. */
429 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
432 memcg = root_mem_cgroup;
433 return &memcg->vmpressure;
436 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
438 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
441 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
443 return (memcg == root_mem_cgroup);
447 * We restrict the id in the range of [1, 65535], so it can fit into
450 #define MEM_CGROUP_ID_MAX USHRT_MAX
452 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
454 return memcg->css.id;
457 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
459 struct cgroup_subsys_state *css;
461 css = css_from_id(id, &memory_cgrp_subsys);
462 return mem_cgroup_from_css(css);
465 /* Writing them here to avoid exposing memcg's inner layout */
466 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
468 void sock_update_memcg(struct sock *sk)
470 if (mem_cgroup_sockets_enabled) {
471 struct mem_cgroup *memcg;
472 struct cg_proto *cg_proto;
474 BUG_ON(!sk->sk_prot->proto_cgroup);
476 /* Socket cloning can throw us here with sk_cgrp already
477 * filled. It won't however, necessarily happen from
478 * process context. So the test for root memcg given
479 * the current task's memcg won't help us in this case.
481 * Respecting the original socket's memcg is a better
482 * decision in this case.
485 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
486 css_get(&sk->sk_cgrp->memcg->css);
491 memcg = mem_cgroup_from_task(current);
492 cg_proto = sk->sk_prot->proto_cgroup(memcg);
493 if (!mem_cgroup_is_root(memcg) &&
494 memcg_proto_active(cg_proto) &&
495 css_tryget_online(&memcg->css)) {
496 sk->sk_cgrp = cg_proto;
501 EXPORT_SYMBOL(sock_update_memcg);
503 void sock_release_memcg(struct sock *sk)
505 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
506 struct mem_cgroup *memcg;
507 WARN_ON(!sk->sk_cgrp->memcg);
508 memcg = sk->sk_cgrp->memcg;
509 css_put(&sk->sk_cgrp->memcg->css);
513 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
515 if (!memcg || mem_cgroup_is_root(memcg))
518 return &memcg->tcp_mem;
520 EXPORT_SYMBOL(tcp_proto_cgroup);
522 static void disarm_sock_keys(struct mem_cgroup *memcg)
524 if (!memcg_proto_activated(&memcg->tcp_mem))
526 static_key_slow_dec(&memcg_socket_limit_enabled);
529 static void disarm_sock_keys(struct mem_cgroup *memcg)
534 #ifdef CONFIG_MEMCG_KMEM
536 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
537 * The main reason for not using cgroup id for this:
538 * this works better in sparse environments, where we have a lot of memcgs,
539 * but only a few kmem-limited. Or also, if we have, for instance, 200
540 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
541 * 200 entry array for that.
543 * The current size of the caches array is stored in memcg_nr_cache_ids. It
544 * will double each time we have to increase it.
546 static DEFINE_IDA(memcg_cache_ida);
547 int memcg_nr_cache_ids;
549 /* Protects memcg_nr_cache_ids */
550 static DECLARE_RWSEM(memcg_cache_ids_sem);
552 void memcg_get_cache_ids(void)
554 down_read(&memcg_cache_ids_sem);
557 void memcg_put_cache_ids(void)
559 up_read(&memcg_cache_ids_sem);
563 * MIN_SIZE is different than 1, because we would like to avoid going through
564 * the alloc/free process all the time. In a small machine, 4 kmem-limited
565 * cgroups is a reasonable guess. In the future, it could be a parameter or
566 * tunable, but that is strictly not necessary.
568 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
569 * this constant directly from cgroup, but it is understandable that this is
570 * better kept as an internal representation in cgroup.c. In any case, the
571 * cgrp_id space is not getting any smaller, and we don't have to necessarily
572 * increase ours as well if it increases.
574 #define MEMCG_CACHES_MIN_SIZE 4
575 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
578 * A lot of the calls to the cache allocation functions are expected to be
579 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
580 * conditional to this static branch, we'll have to allow modules that does
581 * kmem_cache_alloc and the such to see this symbol as well
583 struct static_key memcg_kmem_enabled_key;
584 EXPORT_SYMBOL(memcg_kmem_enabled_key);
586 static void disarm_kmem_keys(struct mem_cgroup *memcg)
588 if (memcg->kmem_acct_activated)
589 static_key_slow_dec(&memcg_kmem_enabled_key);
591 * This check can't live in kmem destruction function,
592 * since the charges will outlive the cgroup
594 WARN_ON(page_counter_read(&memcg->kmem));
597 static void disarm_kmem_keys(struct mem_cgroup *memcg)
600 #endif /* CONFIG_MEMCG_KMEM */
602 static void disarm_static_keys(struct mem_cgroup *memcg)
604 disarm_sock_keys(memcg);
605 disarm_kmem_keys(memcg);
608 static struct mem_cgroup_per_zone *
609 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
611 int nid = zone_to_nid(zone);
612 int zid = zone_idx(zone);
614 return &memcg->nodeinfo[nid]->zoneinfo[zid];
617 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
622 static struct mem_cgroup_per_zone *
623 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
625 int nid = page_to_nid(page);
626 int zid = page_zonenum(page);
628 return &memcg->nodeinfo[nid]->zoneinfo[zid];
631 static struct mem_cgroup_tree_per_zone *
632 soft_limit_tree_node_zone(int nid, int zid)
634 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
637 static struct mem_cgroup_tree_per_zone *
638 soft_limit_tree_from_page(struct page *page)
640 int nid = page_to_nid(page);
641 int zid = page_zonenum(page);
643 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
646 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
647 struct mem_cgroup_tree_per_zone *mctz,
648 unsigned long new_usage_in_excess)
650 struct rb_node **p = &mctz->rb_root.rb_node;
651 struct rb_node *parent = NULL;
652 struct mem_cgroup_per_zone *mz_node;
657 mz->usage_in_excess = new_usage_in_excess;
658 if (!mz->usage_in_excess)
662 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
664 if (mz->usage_in_excess < mz_node->usage_in_excess)
667 * We can't avoid mem cgroups that are over their soft
668 * limit by the same amount
670 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
673 rb_link_node(&mz->tree_node, parent, p);
674 rb_insert_color(&mz->tree_node, &mctz->rb_root);
678 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
679 struct mem_cgroup_tree_per_zone *mctz)
683 rb_erase(&mz->tree_node, &mctz->rb_root);
687 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
688 struct mem_cgroup_tree_per_zone *mctz)
692 spin_lock_irqsave(&mctz->lock, flags);
693 __mem_cgroup_remove_exceeded(mz, mctz);
694 spin_unlock_irqrestore(&mctz->lock, flags);
697 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
699 unsigned long nr_pages = page_counter_read(&memcg->memory);
700 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
701 unsigned long excess = 0;
703 if (nr_pages > soft_limit)
704 excess = nr_pages - soft_limit;
709 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
711 unsigned long excess;
712 struct mem_cgroup_per_zone *mz;
713 struct mem_cgroup_tree_per_zone *mctz;
715 mctz = soft_limit_tree_from_page(page);
717 * Necessary to update all ancestors when hierarchy is used.
718 * because their event counter is not touched.
720 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
721 mz = mem_cgroup_page_zoneinfo(memcg, page);
722 excess = soft_limit_excess(memcg);
724 * We have to update the tree if mz is on RB-tree or
725 * mem is over its softlimit.
727 if (excess || mz->on_tree) {
730 spin_lock_irqsave(&mctz->lock, flags);
731 /* if on-tree, remove it */
733 __mem_cgroup_remove_exceeded(mz, mctz);
735 * Insert again. mz->usage_in_excess will be updated.
736 * If excess is 0, no tree ops.
738 __mem_cgroup_insert_exceeded(mz, mctz, excess);
739 spin_unlock_irqrestore(&mctz->lock, flags);
744 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
746 struct mem_cgroup_tree_per_zone *mctz;
747 struct mem_cgroup_per_zone *mz;
751 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
752 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
753 mctz = soft_limit_tree_node_zone(nid, zid);
754 mem_cgroup_remove_exceeded(mz, mctz);
759 static struct mem_cgroup_per_zone *
760 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
762 struct rb_node *rightmost = NULL;
763 struct mem_cgroup_per_zone *mz;
767 rightmost = rb_last(&mctz->rb_root);
769 goto done; /* Nothing to reclaim from */
771 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
773 * Remove the node now but someone else can add it back,
774 * we will to add it back at the end of reclaim to its correct
775 * position in the tree.
777 __mem_cgroup_remove_exceeded(mz, mctz);
778 if (!soft_limit_excess(mz->memcg) ||
779 !css_tryget_online(&mz->memcg->css))
785 static struct mem_cgroup_per_zone *
786 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
788 struct mem_cgroup_per_zone *mz;
790 spin_lock_irq(&mctz->lock);
791 mz = __mem_cgroup_largest_soft_limit_node(mctz);
792 spin_unlock_irq(&mctz->lock);
797 * Implementation Note: reading percpu statistics for memcg.
799 * Both of vmstat[] and percpu_counter has threshold and do periodic
800 * synchronization to implement "quick" read. There are trade-off between
801 * reading cost and precision of value. Then, we may have a chance to implement
802 * a periodic synchronizion of counter in memcg's counter.
804 * But this _read() function is used for user interface now. The user accounts
805 * memory usage by memory cgroup and he _always_ requires exact value because
806 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
807 * have to visit all online cpus and make sum. So, for now, unnecessary
808 * synchronization is not implemented. (just implemented for cpu hotplug)
810 * If there are kernel internal actions which can make use of some not-exact
811 * value, and reading all cpu value can be performance bottleneck in some
812 * common workload, threashold and synchonization as vmstat[] should be
815 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
816 enum mem_cgroup_stat_index idx)
822 for_each_online_cpu(cpu)
823 val += per_cpu(memcg->stat->count[idx], cpu);
824 #ifdef CONFIG_HOTPLUG_CPU
825 spin_lock(&memcg->pcp_counter_lock);
826 val += memcg->nocpu_base.count[idx];
827 spin_unlock(&memcg->pcp_counter_lock);
833 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
834 enum mem_cgroup_events_index idx)
836 unsigned long val = 0;
840 for_each_online_cpu(cpu)
841 val += per_cpu(memcg->stat->events[idx], cpu);
842 #ifdef CONFIG_HOTPLUG_CPU
843 spin_lock(&memcg->pcp_counter_lock);
844 val += memcg->nocpu_base.events[idx];
845 spin_unlock(&memcg->pcp_counter_lock);
851 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
856 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
857 * counted as CACHE even if it's on ANON LRU.
860 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
863 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
866 if (PageTransHuge(page))
867 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
870 /* pagein of a big page is an event. So, ignore page size */
872 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
874 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
875 nr_pages = -nr_pages; /* for event */
878 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
881 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
883 struct mem_cgroup_per_zone *mz;
885 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
886 return mz->lru_size[lru];
889 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
891 unsigned int lru_mask)
893 unsigned long nr = 0;
896 VM_BUG_ON((unsigned)nid >= nr_node_ids);
898 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
899 struct mem_cgroup_per_zone *mz;
903 if (!(BIT(lru) & lru_mask))
905 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
906 nr += mz->lru_size[lru];
912 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
913 unsigned int lru_mask)
915 unsigned long nr = 0;
918 for_each_node_state(nid, N_MEMORY)
919 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
923 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
924 enum mem_cgroup_events_target target)
926 unsigned long val, next;
928 val = __this_cpu_read(memcg->stat->nr_page_events);
929 next = __this_cpu_read(memcg->stat->targets[target]);
930 /* from time_after() in jiffies.h */
931 if ((long)next - (long)val < 0) {
933 case MEM_CGROUP_TARGET_THRESH:
934 next = val + THRESHOLDS_EVENTS_TARGET;
936 case MEM_CGROUP_TARGET_SOFTLIMIT:
937 next = val + SOFTLIMIT_EVENTS_TARGET;
939 case MEM_CGROUP_TARGET_NUMAINFO:
940 next = val + NUMAINFO_EVENTS_TARGET;
945 __this_cpu_write(memcg->stat->targets[target], next);
952 * Check events in order.
955 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
957 /* threshold event is triggered in finer grain than soft limit */
958 if (unlikely(mem_cgroup_event_ratelimit(memcg,
959 MEM_CGROUP_TARGET_THRESH))) {
961 bool do_numainfo __maybe_unused;
963 do_softlimit = mem_cgroup_event_ratelimit(memcg,
964 MEM_CGROUP_TARGET_SOFTLIMIT);
966 do_numainfo = mem_cgroup_event_ratelimit(memcg,
967 MEM_CGROUP_TARGET_NUMAINFO);
969 mem_cgroup_threshold(memcg);
970 if (unlikely(do_softlimit))
971 mem_cgroup_update_tree(memcg, page);
973 if (unlikely(do_numainfo))
974 atomic_inc(&memcg->numainfo_events);
979 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
982 * mm_update_next_owner() may clear mm->owner to NULL
983 * if it races with swapoff, page migration, etc.
984 * So this can be called with p == NULL.
989 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
992 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
994 struct mem_cgroup *memcg = NULL;
999 * Page cache insertions can happen withou an
1000 * actual mm context, e.g. during disk probing
1001 * on boot, loopback IO, acct() writes etc.
1004 memcg = root_mem_cgroup;
1006 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1007 if (unlikely(!memcg))
1008 memcg = root_mem_cgroup;
1010 } while (!css_tryget_online(&memcg->css));
1016 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1017 * @root: hierarchy root
1018 * @prev: previously returned memcg, NULL on first invocation
1019 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1021 * Returns references to children of the hierarchy below @root, or
1022 * @root itself, or %NULL after a full round-trip.
1024 * Caller must pass the return value in @prev on subsequent
1025 * invocations for reference counting, or use mem_cgroup_iter_break()
1026 * to cancel a hierarchy walk before the round-trip is complete.
1028 * Reclaimers can specify a zone and a priority level in @reclaim to
1029 * divide up the memcgs in the hierarchy among all concurrent
1030 * reclaimers operating on the same zone and priority.
1032 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1033 struct mem_cgroup *prev,
1034 struct mem_cgroup_reclaim_cookie *reclaim)
1036 struct reclaim_iter *uninitialized_var(iter);
1037 struct cgroup_subsys_state *css = NULL;
1038 struct mem_cgroup *memcg = NULL;
1039 struct mem_cgroup *pos = NULL;
1041 if (mem_cgroup_disabled())
1045 root = root_mem_cgroup;
1047 if (prev && !reclaim)
1050 if (!root->use_hierarchy && root != root_mem_cgroup) {
1059 struct mem_cgroup_per_zone *mz;
1061 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1062 iter = &mz->iter[reclaim->priority];
1064 if (prev && reclaim->generation != iter->generation)
1068 pos = ACCESS_ONCE(iter->position);
1070 * A racing update may change the position and
1071 * put the last reference, hence css_tryget(),
1072 * or retry to see the updated position.
1074 } while (pos && !css_tryget(&pos->css));
1081 css = css_next_descendant_pre(css, &root->css);
1084 * Reclaimers share the hierarchy walk, and a
1085 * new one might jump in right at the end of
1086 * the hierarchy - make sure they see at least
1087 * one group and restart from the beginning.
1095 * Verify the css and acquire a reference. The root
1096 * is provided by the caller, so we know it's alive
1097 * and kicking, and don't take an extra reference.
1099 memcg = mem_cgroup_from_css(css);
1101 if (css == &root->css)
1104 if (css_tryget(css)) {
1106 * Make sure the memcg is initialized:
1107 * mem_cgroup_css_online() orders the the
1108 * initialization against setting the flag.
1110 if (smp_load_acquire(&memcg->initialized))
1120 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1122 css_get(&memcg->css);
1128 * pairs with css_tryget when dereferencing iter->position
1137 reclaim->generation = iter->generation;
1143 if (prev && prev != root)
1144 css_put(&prev->css);
1150 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1151 * @root: hierarchy root
1152 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1154 void mem_cgroup_iter_break(struct mem_cgroup *root,
1155 struct mem_cgroup *prev)
1158 root = root_mem_cgroup;
1159 if (prev && prev != root)
1160 css_put(&prev->css);
1164 * Iteration constructs for visiting all cgroups (under a tree). If
1165 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1166 * be used for reference counting.
1168 #define for_each_mem_cgroup_tree(iter, root) \
1169 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1171 iter = mem_cgroup_iter(root, iter, NULL))
1173 #define for_each_mem_cgroup(iter) \
1174 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1176 iter = mem_cgroup_iter(NULL, iter, NULL))
1178 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1180 struct mem_cgroup *memcg;
1183 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1184 if (unlikely(!memcg))
1189 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1192 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1200 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1203 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1204 * @zone: zone of the wanted lruvec
1205 * @memcg: memcg of the wanted lruvec
1207 * Returns the lru list vector holding pages for the given @zone and
1208 * @mem. This can be the global zone lruvec, if the memory controller
1211 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1212 struct mem_cgroup *memcg)
1214 struct mem_cgroup_per_zone *mz;
1215 struct lruvec *lruvec;
1217 if (mem_cgroup_disabled()) {
1218 lruvec = &zone->lruvec;
1222 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1223 lruvec = &mz->lruvec;
1226 * Since a node can be onlined after the mem_cgroup was created,
1227 * we have to be prepared to initialize lruvec->zone here;
1228 * and if offlined then reonlined, we need to reinitialize it.
1230 if (unlikely(lruvec->zone != zone))
1231 lruvec->zone = zone;
1236 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1238 * @zone: zone of the page
1240 * This function is only safe when following the LRU page isolation
1241 * and putback protocol: the LRU lock must be held, and the page must
1242 * either be PageLRU() or the caller must have isolated/allocated it.
1244 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1246 struct mem_cgroup_per_zone *mz;
1247 struct mem_cgroup *memcg;
1248 struct lruvec *lruvec;
1250 if (mem_cgroup_disabled()) {
1251 lruvec = &zone->lruvec;
1255 memcg = page->mem_cgroup;
1257 * Swapcache readahead pages are added to the LRU - and
1258 * possibly migrated - before they are charged.
1261 memcg = root_mem_cgroup;
1263 mz = mem_cgroup_page_zoneinfo(memcg, page);
1264 lruvec = &mz->lruvec;
1267 * Since a node can be onlined after the mem_cgroup was created,
1268 * we have to be prepared to initialize lruvec->zone here;
1269 * and if offlined then reonlined, we need to reinitialize it.
1271 if (unlikely(lruvec->zone != zone))
1272 lruvec->zone = zone;
1277 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1278 * @lruvec: mem_cgroup per zone lru vector
1279 * @lru: index of lru list the page is sitting on
1280 * @nr_pages: positive when adding or negative when removing
1282 * This function must be called when a page is added to or removed from an
1285 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1288 struct mem_cgroup_per_zone *mz;
1289 unsigned long *lru_size;
1291 if (mem_cgroup_disabled())
1294 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1295 lru_size = mz->lru_size + lru;
1296 *lru_size += nr_pages;
1297 VM_BUG_ON((long)(*lru_size) < 0);
1300 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1304 if (!root->use_hierarchy)
1306 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1309 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1311 struct mem_cgroup *task_memcg;
1312 struct task_struct *p;
1315 p = find_lock_task_mm(task);
1317 task_memcg = get_mem_cgroup_from_mm(p->mm);
1321 * All threads may have already detached their mm's, but the oom
1322 * killer still needs to detect if they have already been oom
1323 * killed to prevent needlessly killing additional tasks.
1326 task_memcg = mem_cgroup_from_task(task);
1327 css_get(&task_memcg->css);
1330 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1331 css_put(&task_memcg->css);
1335 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1337 unsigned long inactive_ratio;
1338 unsigned long inactive;
1339 unsigned long active;
1342 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1343 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1345 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1347 inactive_ratio = int_sqrt(10 * gb);
1351 return inactive * inactive_ratio < active;
1354 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1356 struct mem_cgroup_per_zone *mz;
1357 struct mem_cgroup *memcg;
1359 if (mem_cgroup_disabled())
1362 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1365 return !!(memcg->css.flags & CSS_ONLINE);
1368 #define mem_cgroup_from_counter(counter, member) \
1369 container_of(counter, struct mem_cgroup, member)
1372 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1373 * @memcg: the memory cgroup
1375 * Returns the maximum amount of memory @mem can be charged with, in
1378 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1380 unsigned long margin = 0;
1381 unsigned long count;
1382 unsigned long limit;
1384 count = page_counter_read(&memcg->memory);
1385 limit = ACCESS_ONCE(memcg->memory.limit);
1387 margin = limit - count;
1389 if (do_swap_account) {
1390 count = page_counter_read(&memcg->memsw);
1391 limit = ACCESS_ONCE(memcg->memsw.limit);
1393 margin = min(margin, limit - count);
1399 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1402 if (mem_cgroup_disabled() || !memcg->css.parent)
1403 return vm_swappiness;
1405 return memcg->swappiness;
1409 * A routine for checking "mem" is under move_account() or not.
1411 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1412 * moving cgroups. This is for waiting at high-memory pressure
1415 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1417 struct mem_cgroup *from;
1418 struct mem_cgroup *to;
1421 * Unlike task_move routines, we access mc.to, mc.from not under
1422 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1424 spin_lock(&mc.lock);
1430 ret = mem_cgroup_is_descendant(from, memcg) ||
1431 mem_cgroup_is_descendant(to, memcg);
1433 spin_unlock(&mc.lock);
1437 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1439 if (mc.moving_task && current != mc.moving_task) {
1440 if (mem_cgroup_under_move(memcg)) {
1442 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1443 /* moving charge context might have finished. */
1446 finish_wait(&mc.waitq, &wait);
1453 #define K(x) ((x) << (PAGE_SHIFT-10))
1455 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1456 * @memcg: The memory cgroup that went over limit
1457 * @p: Task that is going to be killed
1459 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1462 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1464 /* oom_info_lock ensures that parallel ooms do not interleave */
1465 static DEFINE_MUTEX(oom_info_lock);
1466 struct mem_cgroup *iter;
1472 mutex_lock(&oom_info_lock);
1475 pr_info("Task in ");
1476 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1477 pr_cont(" killed as a result of limit of ");
1478 pr_cont_cgroup_path(memcg->css.cgroup);
1483 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1484 K((u64)page_counter_read(&memcg->memory)),
1485 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1486 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1487 K((u64)page_counter_read(&memcg->memsw)),
1488 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1489 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64)page_counter_read(&memcg->kmem)),
1491 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1493 for_each_mem_cgroup_tree(iter, memcg) {
1494 pr_info("Memory cgroup stats for ");
1495 pr_cont_cgroup_path(iter->css.cgroup);
1498 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1499 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1501 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1502 K(mem_cgroup_read_stat(iter, i)));
1505 for (i = 0; i < NR_LRU_LISTS; i++)
1506 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1507 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1511 mutex_unlock(&oom_info_lock);
1515 * This function returns the number of memcg under hierarchy tree. Returns
1516 * 1(self count) if no children.
1518 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1521 struct mem_cgroup *iter;
1523 for_each_mem_cgroup_tree(iter, memcg)
1529 * Return the memory (and swap, if configured) limit for a memcg.
1531 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1533 unsigned long limit;
1535 limit = memcg->memory.limit;
1536 if (mem_cgroup_swappiness(memcg)) {
1537 unsigned long memsw_limit;
1539 memsw_limit = memcg->memsw.limit;
1540 limit = min(limit + total_swap_pages, memsw_limit);
1545 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1548 struct mem_cgroup *iter;
1549 unsigned long chosen_points = 0;
1550 unsigned long totalpages;
1551 unsigned int points = 0;
1552 struct task_struct *chosen = NULL;
1555 * If current has a pending SIGKILL or is exiting, then automatically
1556 * select it. The goal is to allow it to allocate so that it may
1557 * quickly exit and free its memory.
1559 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1560 mark_tsk_oom_victim(current);
1564 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1565 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1566 for_each_mem_cgroup_tree(iter, memcg) {
1567 struct css_task_iter it;
1568 struct task_struct *task;
1570 css_task_iter_start(&iter->css, &it);
1571 while ((task = css_task_iter_next(&it))) {
1572 switch (oom_scan_process_thread(task, totalpages, NULL,
1574 case OOM_SCAN_SELECT:
1576 put_task_struct(chosen);
1578 chosen_points = ULONG_MAX;
1579 get_task_struct(chosen);
1581 case OOM_SCAN_CONTINUE:
1583 case OOM_SCAN_ABORT:
1584 css_task_iter_end(&it);
1585 mem_cgroup_iter_break(memcg, iter);
1587 put_task_struct(chosen);
1592 points = oom_badness(task, memcg, NULL, totalpages);
1593 if (!points || points < chosen_points)
1595 /* Prefer thread group leaders for display purposes */
1596 if (points == chosen_points &&
1597 thread_group_leader(chosen))
1601 put_task_struct(chosen);
1603 chosen_points = points;
1604 get_task_struct(chosen);
1606 css_task_iter_end(&it);
1611 points = chosen_points * 1000 / totalpages;
1612 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1613 NULL, "Memory cgroup out of memory");
1616 #if MAX_NUMNODES > 1
1619 * test_mem_cgroup_node_reclaimable
1620 * @memcg: the target memcg
1621 * @nid: the node ID to be checked.
1622 * @noswap : specify true here if the user wants flle only information.
1624 * This function returns whether the specified memcg contains any
1625 * reclaimable pages on a node. Returns true if there are any reclaimable
1626 * pages in the node.
1628 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1629 int nid, bool noswap)
1631 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1633 if (noswap || !total_swap_pages)
1635 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1642 * Always updating the nodemask is not very good - even if we have an empty
1643 * list or the wrong list here, we can start from some node and traverse all
1644 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1647 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1651 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1652 * pagein/pageout changes since the last update.
1654 if (!atomic_read(&memcg->numainfo_events))
1656 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1659 /* make a nodemask where this memcg uses memory from */
1660 memcg->scan_nodes = node_states[N_MEMORY];
1662 for_each_node_mask(nid, node_states[N_MEMORY]) {
1664 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1665 node_clear(nid, memcg->scan_nodes);
1668 atomic_set(&memcg->numainfo_events, 0);
1669 atomic_set(&memcg->numainfo_updating, 0);
1673 * Selecting a node where we start reclaim from. Because what we need is just
1674 * reducing usage counter, start from anywhere is O,K. Considering
1675 * memory reclaim from current node, there are pros. and cons.
1677 * Freeing memory from current node means freeing memory from a node which
1678 * we'll use or we've used. So, it may make LRU bad. And if several threads
1679 * hit limits, it will see a contention on a node. But freeing from remote
1680 * node means more costs for memory reclaim because of memory latency.
1682 * Now, we use round-robin. Better algorithm is welcomed.
1684 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1688 mem_cgroup_may_update_nodemask(memcg);
1689 node = memcg->last_scanned_node;
1691 node = next_node(node, memcg->scan_nodes);
1692 if (node == MAX_NUMNODES)
1693 node = first_node(memcg->scan_nodes);
1695 * We call this when we hit limit, not when pages are added to LRU.
1696 * No LRU may hold pages because all pages are UNEVICTABLE or
1697 * memcg is too small and all pages are not on LRU. In that case,
1698 * we use curret node.
1700 if (unlikely(node == MAX_NUMNODES))
1701 node = numa_node_id();
1703 memcg->last_scanned_node = node;
1707 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1713 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1716 unsigned long *total_scanned)
1718 struct mem_cgroup *victim = NULL;
1721 unsigned long excess;
1722 unsigned long nr_scanned;
1723 struct mem_cgroup_reclaim_cookie reclaim = {
1728 excess = soft_limit_excess(root_memcg);
1731 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1736 * If we have not been able to reclaim
1737 * anything, it might because there are
1738 * no reclaimable pages under this hierarchy
1743 * We want to do more targeted reclaim.
1744 * excess >> 2 is not to excessive so as to
1745 * reclaim too much, nor too less that we keep
1746 * coming back to reclaim from this cgroup
1748 if (total >= (excess >> 2) ||
1749 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1754 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1756 *total_scanned += nr_scanned;
1757 if (!soft_limit_excess(root_memcg))
1760 mem_cgroup_iter_break(root_memcg, victim);
1764 #ifdef CONFIG_LOCKDEP
1765 static struct lockdep_map memcg_oom_lock_dep_map = {
1766 .name = "memcg_oom_lock",
1770 static DEFINE_SPINLOCK(memcg_oom_lock);
1773 * Check OOM-Killer is already running under our hierarchy.
1774 * If someone is running, return false.
1776 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter, *failed = NULL;
1780 spin_lock(&memcg_oom_lock);
1782 for_each_mem_cgroup_tree(iter, memcg) {
1783 if (iter->oom_lock) {
1785 * this subtree of our hierarchy is already locked
1786 * so we cannot give a lock.
1789 mem_cgroup_iter_break(memcg, iter);
1792 iter->oom_lock = true;
1797 * OK, we failed to lock the whole subtree so we have
1798 * to clean up what we set up to the failing subtree
1800 for_each_mem_cgroup_tree(iter, memcg) {
1801 if (iter == failed) {
1802 mem_cgroup_iter_break(memcg, iter);
1805 iter->oom_lock = false;
1808 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1810 spin_unlock(&memcg_oom_lock);
1815 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1819 spin_lock(&memcg_oom_lock);
1820 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1821 for_each_mem_cgroup_tree(iter, memcg)
1822 iter->oom_lock = false;
1823 spin_unlock(&memcg_oom_lock);
1826 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1828 struct mem_cgroup *iter;
1830 for_each_mem_cgroup_tree(iter, memcg)
1831 atomic_inc(&iter->under_oom);
1834 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1836 struct mem_cgroup *iter;
1839 * When a new child is created while the hierarchy is under oom,
1840 * mem_cgroup_oom_lock() may not be called. We have to use
1841 * atomic_add_unless() here.
1843 for_each_mem_cgroup_tree(iter, memcg)
1844 atomic_add_unless(&iter->under_oom, -1, 0);
1847 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1849 struct oom_wait_info {
1850 struct mem_cgroup *memcg;
1854 static int memcg_oom_wake_function(wait_queue_t *wait,
1855 unsigned mode, int sync, void *arg)
1857 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1858 struct mem_cgroup *oom_wait_memcg;
1859 struct oom_wait_info *oom_wait_info;
1861 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1862 oom_wait_memcg = oom_wait_info->memcg;
1864 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1865 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1867 return autoremove_wake_function(wait, mode, sync, arg);
1870 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1872 atomic_inc(&memcg->oom_wakeups);
1873 /* for filtering, pass "memcg" as argument. */
1874 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1877 static void memcg_oom_recover(struct mem_cgroup *memcg)
1879 if (memcg && atomic_read(&memcg->under_oom))
1880 memcg_wakeup_oom(memcg);
1883 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1885 if (!current->memcg_oom.may_oom)
1888 * We are in the middle of the charge context here, so we
1889 * don't want to block when potentially sitting on a callstack
1890 * that holds all kinds of filesystem and mm locks.
1892 * Also, the caller may handle a failed allocation gracefully
1893 * (like optional page cache readahead) and so an OOM killer
1894 * invocation might not even be necessary.
1896 * That's why we don't do anything here except remember the
1897 * OOM context and then deal with it at the end of the page
1898 * fault when the stack is unwound, the locks are released,
1899 * and when we know whether the fault was overall successful.
1901 css_get(&memcg->css);
1902 current->memcg_oom.memcg = memcg;
1903 current->memcg_oom.gfp_mask = mask;
1904 current->memcg_oom.order = order;
1908 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1909 * @handle: actually kill/wait or just clean up the OOM state
1911 * This has to be called at the end of a page fault if the memcg OOM
1912 * handler was enabled.
1914 * Memcg supports userspace OOM handling where failed allocations must
1915 * sleep on a waitqueue until the userspace task resolves the
1916 * situation. Sleeping directly in the charge context with all kinds
1917 * of locks held is not a good idea, instead we remember an OOM state
1918 * in the task and mem_cgroup_oom_synchronize() has to be called at
1919 * the end of the page fault to complete the OOM handling.
1921 * Returns %true if an ongoing memcg OOM situation was detected and
1922 * completed, %false otherwise.
1924 bool mem_cgroup_oom_synchronize(bool handle)
1926 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1927 struct oom_wait_info owait;
1930 /* OOM is global, do not handle */
1934 if (!handle || oom_killer_disabled)
1937 owait.memcg = memcg;
1938 owait.wait.flags = 0;
1939 owait.wait.func = memcg_oom_wake_function;
1940 owait.wait.private = current;
1941 INIT_LIST_HEAD(&owait.wait.task_list);
1943 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1944 mem_cgroup_mark_under_oom(memcg);
1946 locked = mem_cgroup_oom_trylock(memcg);
1949 mem_cgroup_oom_notify(memcg);
1951 if (locked && !memcg->oom_kill_disable) {
1952 mem_cgroup_unmark_under_oom(memcg);
1953 finish_wait(&memcg_oom_waitq, &owait.wait);
1954 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1955 current->memcg_oom.order);
1958 mem_cgroup_unmark_under_oom(memcg);
1959 finish_wait(&memcg_oom_waitq, &owait.wait);
1963 mem_cgroup_oom_unlock(memcg);
1965 * There is no guarantee that an OOM-lock contender
1966 * sees the wakeups triggered by the OOM kill
1967 * uncharges. Wake any sleepers explicitely.
1969 memcg_oom_recover(memcg);
1972 current->memcg_oom.memcg = NULL;
1973 css_put(&memcg->css);
1978 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1979 * @page: page that is going to change accounted state
1981 * This function must mark the beginning of an accounted page state
1982 * change to prevent double accounting when the page is concurrently
1983 * being moved to another memcg:
1985 * memcg = mem_cgroup_begin_page_stat(page);
1986 * if (TestClearPageState(page))
1987 * mem_cgroup_update_page_stat(memcg, state, -1);
1988 * mem_cgroup_end_page_stat(memcg);
1990 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1992 struct mem_cgroup *memcg;
1993 unsigned long flags;
1996 * The RCU lock is held throughout the transaction. The fast
1997 * path can get away without acquiring the memcg->move_lock
1998 * because page moving starts with an RCU grace period.
2000 * The RCU lock also protects the memcg from being freed when
2001 * the page state that is going to change is the only thing
2002 * preventing the page from being uncharged.
2003 * E.g. end-writeback clearing PageWriteback(), which allows
2004 * migration to go ahead and uncharge the page before the
2005 * account transaction might be complete.
2009 if (mem_cgroup_disabled())
2012 memcg = page->mem_cgroup;
2013 if (unlikely(!memcg))
2016 if (atomic_read(&memcg->moving_account) <= 0)
2019 spin_lock_irqsave(&memcg->move_lock, flags);
2020 if (memcg != page->mem_cgroup) {
2021 spin_unlock_irqrestore(&memcg->move_lock, flags);
2026 * When charge migration first begins, we can have locked and
2027 * unlocked page stat updates happening concurrently. Track
2028 * the task who has the lock for mem_cgroup_end_page_stat().
2030 memcg->move_lock_task = current;
2031 memcg->move_lock_flags = flags;
2037 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2038 * @memcg: the memcg that was accounted against
2040 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2042 if (memcg && memcg->move_lock_task == current) {
2043 unsigned long flags = memcg->move_lock_flags;
2045 memcg->move_lock_task = NULL;
2046 memcg->move_lock_flags = 0;
2048 spin_unlock_irqrestore(&memcg->move_lock, flags);
2055 * mem_cgroup_update_page_stat - update page state statistics
2056 * @memcg: memcg to account against
2057 * @idx: page state item to account
2058 * @val: number of pages (positive or negative)
2060 * See mem_cgroup_begin_page_stat() for locking requirements.
2062 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2063 enum mem_cgroup_stat_index idx, int val)
2065 VM_BUG_ON(!rcu_read_lock_held());
2068 this_cpu_add(memcg->stat->count[idx], val);
2072 * size of first charge trial. "32" comes from vmscan.c's magic value.
2073 * TODO: maybe necessary to use big numbers in big irons.
2075 #define CHARGE_BATCH 32U
2076 struct memcg_stock_pcp {
2077 struct mem_cgroup *cached; /* this never be root cgroup */
2078 unsigned int nr_pages;
2079 struct work_struct work;
2080 unsigned long flags;
2081 #define FLUSHING_CACHED_CHARGE 0
2083 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2084 static DEFINE_MUTEX(percpu_charge_mutex);
2087 * consume_stock: Try to consume stocked charge on this cpu.
2088 * @memcg: memcg to consume from.
2089 * @nr_pages: how many pages to charge.
2091 * The charges will only happen if @memcg matches the current cpu's memcg
2092 * stock, and at least @nr_pages are available in that stock. Failure to
2093 * service an allocation will refill the stock.
2095 * returns true if successful, false otherwise.
2097 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2099 struct memcg_stock_pcp *stock;
2102 if (nr_pages > CHARGE_BATCH)
2105 stock = &get_cpu_var(memcg_stock);
2106 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2107 stock->nr_pages -= nr_pages;
2110 put_cpu_var(memcg_stock);
2115 * Returns stocks cached in percpu and reset cached information.
2117 static void drain_stock(struct memcg_stock_pcp *stock)
2119 struct mem_cgroup *old = stock->cached;
2121 if (stock->nr_pages) {
2122 page_counter_uncharge(&old->memory, stock->nr_pages);
2123 if (do_swap_account)
2124 page_counter_uncharge(&old->memsw, stock->nr_pages);
2125 css_put_many(&old->css, stock->nr_pages);
2126 stock->nr_pages = 0;
2128 stock->cached = NULL;
2132 * This must be called under preempt disabled or must be called by
2133 * a thread which is pinned to local cpu.
2135 static void drain_local_stock(struct work_struct *dummy)
2137 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2139 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2143 * Cache charges(val) to local per_cpu area.
2144 * This will be consumed by consume_stock() function, later.
2146 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2148 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2150 if (stock->cached != memcg) { /* reset if necessary */
2152 stock->cached = memcg;
2154 stock->nr_pages += nr_pages;
2155 put_cpu_var(memcg_stock);
2159 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2160 * of the hierarchy under it.
2162 static void drain_all_stock(struct mem_cgroup *root_memcg)
2166 /* If someone's already draining, avoid adding running more workers. */
2167 if (!mutex_trylock(&percpu_charge_mutex))
2169 /* Notify other cpus that system-wide "drain" is running */
2172 for_each_online_cpu(cpu) {
2173 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2174 struct mem_cgroup *memcg;
2176 memcg = stock->cached;
2177 if (!memcg || !stock->nr_pages)
2179 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2181 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2183 drain_local_stock(&stock->work);
2185 schedule_work_on(cpu, &stock->work);
2190 mutex_unlock(&percpu_charge_mutex);
2194 * This function drains percpu counter value from DEAD cpu and
2195 * move it to local cpu. Note that this function can be preempted.
2197 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2201 spin_lock(&memcg->pcp_counter_lock);
2202 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2203 long x = per_cpu(memcg->stat->count[i], cpu);
2205 per_cpu(memcg->stat->count[i], cpu) = 0;
2206 memcg->nocpu_base.count[i] += x;
2208 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2209 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2211 per_cpu(memcg->stat->events[i], cpu) = 0;
2212 memcg->nocpu_base.events[i] += x;
2214 spin_unlock(&memcg->pcp_counter_lock);
2217 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2218 unsigned long action,
2221 int cpu = (unsigned long)hcpu;
2222 struct memcg_stock_pcp *stock;
2223 struct mem_cgroup *iter;
2225 if (action == CPU_ONLINE)
2228 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2231 for_each_mem_cgroup(iter)
2232 mem_cgroup_drain_pcp_counter(iter, cpu);
2234 stock = &per_cpu(memcg_stock, cpu);
2239 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2240 unsigned int nr_pages)
2242 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2243 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2244 struct mem_cgroup *mem_over_limit;
2245 struct page_counter *counter;
2246 unsigned long nr_reclaimed;
2247 bool may_swap = true;
2248 bool drained = false;
2251 if (mem_cgroup_is_root(memcg))
2254 if (consume_stock(memcg, nr_pages))
2257 if (!do_swap_account ||
2258 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2259 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2261 if (do_swap_account)
2262 page_counter_uncharge(&memcg->memsw, batch);
2263 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2265 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2269 if (batch > nr_pages) {
2275 * Unlike in global OOM situations, memcg is not in a physical
2276 * memory shortage. Allow dying and OOM-killed tasks to
2277 * bypass the last charges so that they can exit quickly and
2278 * free their memory.
2280 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2281 fatal_signal_pending(current) ||
2282 current->flags & PF_EXITING))
2285 if (unlikely(task_in_memcg_oom(current)))
2288 if (!(gfp_mask & __GFP_WAIT))
2291 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2293 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2294 gfp_mask, may_swap);
2296 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2300 drain_all_stock(mem_over_limit);
2305 if (gfp_mask & __GFP_NORETRY)
2308 * Even though the limit is exceeded at this point, reclaim
2309 * may have been able to free some pages. Retry the charge
2310 * before killing the task.
2312 * Only for regular pages, though: huge pages are rather
2313 * unlikely to succeed so close to the limit, and we fall back
2314 * to regular pages anyway in case of failure.
2316 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2319 * At task move, charge accounts can be doubly counted. So, it's
2320 * better to wait until the end of task_move if something is going on.
2322 if (mem_cgroup_wait_acct_move(mem_over_limit))
2328 if (gfp_mask & __GFP_NOFAIL)
2331 if (fatal_signal_pending(current))
2334 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2336 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2338 if (!(gfp_mask & __GFP_NOFAIL))
2344 css_get_many(&memcg->css, batch);
2345 if (batch > nr_pages)
2346 refill_stock(memcg, batch - nr_pages);
2348 * If the hierarchy is above the normal consumption range,
2349 * make the charging task trim their excess contribution.
2352 if (page_counter_read(&memcg->memory) <= memcg->high)
2354 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2355 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2356 } while ((memcg = parent_mem_cgroup(memcg)));
2361 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2363 if (mem_cgroup_is_root(memcg))
2366 page_counter_uncharge(&memcg->memory, nr_pages);
2367 if (do_swap_account)
2368 page_counter_uncharge(&memcg->memsw, nr_pages);
2370 css_put_many(&memcg->css, nr_pages);
2374 * A helper function to get mem_cgroup from ID. must be called under
2375 * rcu_read_lock(). The caller is responsible for calling
2376 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2377 * refcnt from swap can be called against removed memcg.)
2379 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2381 /* ID 0 is unused ID */
2384 return mem_cgroup_from_id(id);
2388 * try_get_mem_cgroup_from_page - look up page's memcg association
2391 * Look up, get a css reference, and return the memcg that owns @page.
2393 * The page must be locked to prevent racing with swap-in and page
2394 * cache charges. If coming from an unlocked page table, the caller
2395 * must ensure the page is on the LRU or this can race with charging.
2397 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2399 struct mem_cgroup *memcg;
2403 VM_BUG_ON_PAGE(!PageLocked(page), page);
2405 memcg = page->mem_cgroup;
2407 if (!css_tryget_online(&memcg->css))
2409 } else if (PageSwapCache(page)) {
2410 ent.val = page_private(page);
2411 id = lookup_swap_cgroup_id(ent);
2413 memcg = mem_cgroup_lookup(id);
2414 if (memcg && !css_tryget_online(&memcg->css))
2421 static void lock_page_lru(struct page *page, int *isolated)
2423 struct zone *zone = page_zone(page);
2425 spin_lock_irq(&zone->lru_lock);
2426 if (PageLRU(page)) {
2427 struct lruvec *lruvec;
2429 lruvec = mem_cgroup_page_lruvec(page, zone);
2431 del_page_from_lru_list(page, lruvec, page_lru(page));
2437 static void unlock_page_lru(struct page *page, int isolated)
2439 struct zone *zone = page_zone(page);
2442 struct lruvec *lruvec;
2444 lruvec = mem_cgroup_page_lruvec(page, zone);
2445 VM_BUG_ON_PAGE(PageLRU(page), page);
2447 add_page_to_lru_list(page, lruvec, page_lru(page));
2449 spin_unlock_irq(&zone->lru_lock);
2452 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2457 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2460 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2461 * may already be on some other mem_cgroup's LRU. Take care of it.
2464 lock_page_lru(page, &isolated);
2467 * Nobody should be changing or seriously looking at
2468 * page->mem_cgroup at this point:
2470 * - the page is uncharged
2472 * - the page is off-LRU
2474 * - an anonymous fault has exclusive page access, except for
2475 * a locked page table
2477 * - a page cache insertion, a swapin fault, or a migration
2478 * have the page locked
2480 page->mem_cgroup = memcg;
2483 unlock_page_lru(page, isolated);
2486 #ifdef CONFIG_MEMCG_KMEM
2487 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2488 unsigned long nr_pages)
2490 struct page_counter *counter;
2493 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2497 ret = try_charge(memcg, gfp, nr_pages);
2498 if (ret == -EINTR) {
2500 * try_charge() chose to bypass to root due to OOM kill or
2501 * fatal signal. Since our only options are to either fail
2502 * the allocation or charge it to this cgroup, do it as a
2503 * temporary condition. But we can't fail. From a kmem/slab
2504 * perspective, the cache has already been selected, by
2505 * mem_cgroup_kmem_get_cache(), so it is too late to change
2508 * This condition will only trigger if the task entered
2509 * memcg_charge_kmem in a sane state, but was OOM-killed
2510 * during try_charge() above. Tasks that were already dying
2511 * when the allocation triggers should have been already
2512 * directed to the root cgroup in memcontrol.h
2514 page_counter_charge(&memcg->memory, nr_pages);
2515 if (do_swap_account)
2516 page_counter_charge(&memcg->memsw, nr_pages);
2517 css_get_many(&memcg->css, nr_pages);
2520 page_counter_uncharge(&memcg->kmem, nr_pages);
2525 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2527 page_counter_uncharge(&memcg->memory, nr_pages);
2528 if (do_swap_account)
2529 page_counter_uncharge(&memcg->memsw, nr_pages);
2531 page_counter_uncharge(&memcg->kmem, nr_pages);
2533 css_put_many(&memcg->css, nr_pages);
2537 * helper for acessing a memcg's index. It will be used as an index in the
2538 * child cache array in kmem_cache, and also to derive its name. This function
2539 * will return -1 when this is not a kmem-limited memcg.
2541 int memcg_cache_id(struct mem_cgroup *memcg)
2543 return memcg ? memcg->kmemcg_id : -1;
2546 static int memcg_alloc_cache_id(void)
2551 id = ida_simple_get(&memcg_cache_ida,
2552 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2556 if (id < memcg_nr_cache_ids)
2560 * There's no space for the new id in memcg_caches arrays,
2561 * so we have to grow them.
2563 down_write(&memcg_cache_ids_sem);
2565 size = 2 * (id + 1);
2566 if (size < MEMCG_CACHES_MIN_SIZE)
2567 size = MEMCG_CACHES_MIN_SIZE;
2568 else if (size > MEMCG_CACHES_MAX_SIZE)
2569 size = MEMCG_CACHES_MAX_SIZE;
2571 err = memcg_update_all_caches(size);
2573 err = memcg_update_all_list_lrus(size);
2575 memcg_nr_cache_ids = size;
2577 up_write(&memcg_cache_ids_sem);
2580 ida_simple_remove(&memcg_cache_ida, id);
2586 static void memcg_free_cache_id(int id)
2588 ida_simple_remove(&memcg_cache_ida, id);
2591 struct memcg_kmem_cache_create_work {
2592 struct mem_cgroup *memcg;
2593 struct kmem_cache *cachep;
2594 struct work_struct work;
2597 static void memcg_kmem_cache_create_func(struct work_struct *w)
2599 struct memcg_kmem_cache_create_work *cw =
2600 container_of(w, struct memcg_kmem_cache_create_work, work);
2601 struct mem_cgroup *memcg = cw->memcg;
2602 struct kmem_cache *cachep = cw->cachep;
2604 memcg_create_kmem_cache(memcg, cachep);
2606 css_put(&memcg->css);
2611 * Enqueue the creation of a per-memcg kmem_cache.
2613 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2614 struct kmem_cache *cachep)
2616 struct memcg_kmem_cache_create_work *cw;
2618 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2622 css_get(&memcg->css);
2625 cw->cachep = cachep;
2626 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2628 schedule_work(&cw->work);
2631 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2632 struct kmem_cache *cachep)
2635 * We need to stop accounting when we kmalloc, because if the
2636 * corresponding kmalloc cache is not yet created, the first allocation
2637 * in __memcg_schedule_kmem_cache_create will recurse.
2639 * However, it is better to enclose the whole function. Depending on
2640 * the debugging options enabled, INIT_WORK(), for instance, can
2641 * trigger an allocation. This too, will make us recurse. Because at
2642 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2643 * the safest choice is to do it like this, wrapping the whole function.
2645 current->memcg_kmem_skip_account = 1;
2646 __memcg_schedule_kmem_cache_create(memcg, cachep);
2647 current->memcg_kmem_skip_account = 0;
2651 * Return the kmem_cache we're supposed to use for a slab allocation.
2652 * We try to use the current memcg's version of the cache.
2654 * If the cache does not exist yet, if we are the first user of it,
2655 * we either create it immediately, if possible, or create it asynchronously
2657 * In the latter case, we will let the current allocation go through with
2658 * the original cache.
2660 * Can't be called in interrupt context or from kernel threads.
2661 * This function needs to be called with rcu_read_lock() held.
2663 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2665 struct mem_cgroup *memcg;
2666 struct kmem_cache *memcg_cachep;
2669 VM_BUG_ON(!is_root_cache(cachep));
2671 if (current->memcg_kmem_skip_account)
2674 memcg = get_mem_cgroup_from_mm(current->mm);
2675 kmemcg_id = ACCESS_ONCE(memcg->kmemcg_id);
2679 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2680 if (likely(memcg_cachep))
2681 return memcg_cachep;
2684 * If we are in a safe context (can wait, and not in interrupt
2685 * context), we could be be predictable and return right away.
2686 * This would guarantee that the allocation being performed
2687 * already belongs in the new cache.
2689 * However, there are some clashes that can arrive from locking.
2690 * For instance, because we acquire the slab_mutex while doing
2691 * memcg_create_kmem_cache, this means no further allocation
2692 * could happen with the slab_mutex held. So it's better to
2695 memcg_schedule_kmem_cache_create(memcg, cachep);
2697 css_put(&memcg->css);
2701 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2703 if (!is_root_cache(cachep))
2704 css_put(&cachep->memcg_params.memcg->css);
2708 * We need to verify if the allocation against current->mm->owner's memcg is
2709 * possible for the given order. But the page is not allocated yet, so we'll
2710 * need a further commit step to do the final arrangements.
2712 * It is possible for the task to switch cgroups in this mean time, so at
2713 * commit time, we can't rely on task conversion any longer. We'll then use
2714 * the handle argument to return to the caller which cgroup we should commit
2715 * against. We could also return the memcg directly and avoid the pointer
2716 * passing, but a boolean return value gives better semantics considering
2717 * the compiled-out case as well.
2719 * Returning true means the allocation is possible.
2722 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2724 struct mem_cgroup *memcg;
2729 memcg = get_mem_cgroup_from_mm(current->mm);
2731 if (!memcg_kmem_is_active(memcg)) {
2732 css_put(&memcg->css);
2736 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2740 css_put(&memcg->css);
2744 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2747 VM_BUG_ON(mem_cgroup_is_root(memcg));
2749 /* The page allocation failed. Revert */
2751 memcg_uncharge_kmem(memcg, 1 << order);
2754 page->mem_cgroup = memcg;
2757 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2759 struct mem_cgroup *memcg = page->mem_cgroup;
2764 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2766 memcg_uncharge_kmem(memcg, 1 << order);
2767 page->mem_cgroup = NULL;
2770 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2772 struct mem_cgroup *memcg = NULL;
2773 struct kmem_cache *cachep;
2776 page = virt_to_head_page(ptr);
2777 if (PageSlab(page)) {
2778 cachep = page->slab_cache;
2779 if (!is_root_cache(cachep))
2780 memcg = cachep->memcg_params.memcg;
2782 /* page allocated by alloc_kmem_pages */
2783 memcg = page->mem_cgroup;
2787 #endif /* CONFIG_MEMCG_KMEM */
2789 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2792 * Because tail pages are not marked as "used", set it. We're under
2793 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2794 * charge/uncharge will be never happen and move_account() is done under
2795 * compound_lock(), so we don't have to take care of races.
2797 void mem_cgroup_split_huge_fixup(struct page *head)
2801 if (mem_cgroup_disabled())
2804 for (i = 1; i < HPAGE_PMD_NR; i++)
2805 head[i].mem_cgroup = head->mem_cgroup;
2807 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2810 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2813 * mem_cgroup_move_account - move account of the page
2815 * @nr_pages: number of regular pages (>1 for huge pages)
2816 * @from: mem_cgroup which the page is moved from.
2817 * @to: mem_cgroup which the page is moved to. @from != @to.
2819 * The caller must confirm following.
2820 * - page is not on LRU (isolate_page() is useful.)
2821 * - compound_lock is held when nr_pages > 1
2823 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2826 static int mem_cgroup_move_account(struct page *page,
2827 unsigned int nr_pages,
2828 struct mem_cgroup *from,
2829 struct mem_cgroup *to)
2831 unsigned long flags;
2834 VM_BUG_ON(from == to);
2835 VM_BUG_ON_PAGE(PageLRU(page), page);
2837 * The page is isolated from LRU. So, collapse function
2838 * will not handle this page. But page splitting can happen.
2839 * Do this check under compound_page_lock(). The caller should
2843 if (nr_pages > 1 && !PageTransHuge(page))
2847 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2848 * of its source page while we change it: page migration takes
2849 * both pages off the LRU, but page cache replacement doesn't.
2851 if (!trylock_page(page))
2855 if (page->mem_cgroup != from)
2858 spin_lock_irqsave(&from->move_lock, flags);
2860 if (!PageAnon(page) && page_mapped(page)) {
2861 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2863 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2867 if (PageWriteback(page)) {
2868 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2870 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2875 * It is safe to change page->mem_cgroup here because the page
2876 * is referenced, charged, and isolated - we can't race with
2877 * uncharging, charging, migration, or LRU putback.
2880 /* caller should have done css_get */
2881 page->mem_cgroup = to;
2882 spin_unlock_irqrestore(&from->move_lock, flags);
2886 local_irq_disable();
2887 mem_cgroup_charge_statistics(to, page, nr_pages);
2888 memcg_check_events(to, page);
2889 mem_cgroup_charge_statistics(from, page, -nr_pages);
2890 memcg_check_events(from, page);
2898 #ifdef CONFIG_MEMCG_SWAP
2899 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2902 int val = (charge) ? 1 : -1;
2903 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2907 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2908 * @entry: swap entry to be moved
2909 * @from: mem_cgroup which the entry is moved from
2910 * @to: mem_cgroup which the entry is moved to
2912 * It succeeds only when the swap_cgroup's record for this entry is the same
2913 * as the mem_cgroup's id of @from.
2915 * Returns 0 on success, -EINVAL on failure.
2917 * The caller must have charged to @to, IOW, called page_counter_charge() about
2918 * both res and memsw, and called css_get().
2920 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2921 struct mem_cgroup *from, struct mem_cgroup *to)
2923 unsigned short old_id, new_id;
2925 old_id = mem_cgroup_id(from);
2926 new_id = mem_cgroup_id(to);
2928 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2929 mem_cgroup_swap_statistics(from, false);
2930 mem_cgroup_swap_statistics(to, true);
2936 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2937 struct mem_cgroup *from, struct mem_cgroup *to)
2943 static DEFINE_MUTEX(memcg_limit_mutex);
2945 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2946 unsigned long limit)
2948 unsigned long curusage;
2949 unsigned long oldusage;
2950 bool enlarge = false;
2955 * For keeping hierarchical_reclaim simple, how long we should retry
2956 * is depends on callers. We set our retry-count to be function
2957 * of # of children which we should visit in this loop.
2959 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2960 mem_cgroup_count_children(memcg);
2962 oldusage = page_counter_read(&memcg->memory);
2965 if (signal_pending(current)) {
2970 mutex_lock(&memcg_limit_mutex);
2971 if (limit > memcg->memsw.limit) {
2972 mutex_unlock(&memcg_limit_mutex);
2976 if (limit > memcg->memory.limit)
2978 ret = page_counter_limit(&memcg->memory, limit);
2979 mutex_unlock(&memcg_limit_mutex);
2984 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2986 curusage = page_counter_read(&memcg->memory);
2987 /* Usage is reduced ? */
2988 if (curusage >= oldusage)
2991 oldusage = curusage;
2992 } while (retry_count);
2994 if (!ret && enlarge)
2995 memcg_oom_recover(memcg);
3000 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3001 unsigned long limit)
3003 unsigned long curusage;
3004 unsigned long oldusage;
3005 bool enlarge = false;
3009 /* see mem_cgroup_resize_res_limit */
3010 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3011 mem_cgroup_count_children(memcg);
3013 oldusage = page_counter_read(&memcg->memsw);
3016 if (signal_pending(current)) {
3021 mutex_lock(&memcg_limit_mutex);
3022 if (limit < memcg->memory.limit) {
3023 mutex_unlock(&memcg_limit_mutex);
3027 if (limit > memcg->memsw.limit)
3029 ret = page_counter_limit(&memcg->memsw, limit);
3030 mutex_unlock(&memcg_limit_mutex);
3035 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3037 curusage = page_counter_read(&memcg->memsw);
3038 /* Usage is reduced ? */
3039 if (curusage >= oldusage)
3042 oldusage = curusage;
3043 } while (retry_count);
3045 if (!ret && enlarge)
3046 memcg_oom_recover(memcg);
3051 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3053 unsigned long *total_scanned)
3055 unsigned long nr_reclaimed = 0;
3056 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3057 unsigned long reclaimed;
3059 struct mem_cgroup_tree_per_zone *mctz;
3060 unsigned long excess;
3061 unsigned long nr_scanned;
3066 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3068 * This loop can run a while, specially if mem_cgroup's continuously
3069 * keep exceeding their soft limit and putting the system under
3076 mz = mem_cgroup_largest_soft_limit_node(mctz);
3081 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3082 gfp_mask, &nr_scanned);
3083 nr_reclaimed += reclaimed;
3084 *total_scanned += nr_scanned;
3085 spin_lock_irq(&mctz->lock);
3086 __mem_cgroup_remove_exceeded(mz, mctz);
3089 * If we failed to reclaim anything from this memory cgroup
3090 * it is time to move on to the next cgroup
3094 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3096 excess = soft_limit_excess(mz->memcg);
3098 * One school of thought says that we should not add
3099 * back the node to the tree if reclaim returns 0.
3100 * But our reclaim could return 0, simply because due
3101 * to priority we are exposing a smaller subset of
3102 * memory to reclaim from. Consider this as a longer
3105 /* If excess == 0, no tree ops */
3106 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3107 spin_unlock_irq(&mctz->lock);
3108 css_put(&mz->memcg->css);
3111 * Could not reclaim anything and there are no more
3112 * mem cgroups to try or we seem to be looping without
3113 * reclaiming anything.
3115 if (!nr_reclaimed &&
3117 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3119 } while (!nr_reclaimed);
3121 css_put(&next_mz->memcg->css);
3122 return nr_reclaimed;
3126 * Test whether @memcg has children, dead or alive. Note that this
3127 * function doesn't care whether @memcg has use_hierarchy enabled and
3128 * returns %true if there are child csses according to the cgroup
3129 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3131 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3136 * The lock does not prevent addition or deletion of children, but
3137 * it prevents a new child from being initialized based on this
3138 * parent in css_online(), so it's enough to decide whether
3139 * hierarchically inherited attributes can still be changed or not.
3141 lockdep_assert_held(&memcg_create_mutex);
3144 ret = css_next_child(NULL, &memcg->css);
3150 * Reclaims as many pages from the given memcg as possible and moves
3151 * the rest to the parent.
3153 * Caller is responsible for holding css reference for memcg.
3155 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3157 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3159 /* we call try-to-free pages for make this cgroup empty */
3160 lru_add_drain_all();
3161 /* try to free all pages in this cgroup */
3162 while (nr_retries && page_counter_read(&memcg->memory)) {
3165 if (signal_pending(current))
3168 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3172 /* maybe some writeback is necessary */
3173 congestion_wait(BLK_RW_ASYNC, HZ/10);
3181 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3182 char *buf, size_t nbytes,
3185 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3187 if (mem_cgroup_is_root(memcg))
3189 return mem_cgroup_force_empty(memcg) ?: nbytes;
3192 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3195 return mem_cgroup_from_css(css)->use_hierarchy;
3198 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3199 struct cftype *cft, u64 val)
3202 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3203 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3205 mutex_lock(&memcg_create_mutex);
3207 if (memcg->use_hierarchy == val)
3211 * If parent's use_hierarchy is set, we can't make any modifications
3212 * in the child subtrees. If it is unset, then the change can
3213 * occur, provided the current cgroup has no children.
3215 * For the root cgroup, parent_mem is NULL, we allow value to be
3216 * set if there are no children.
3218 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3219 (val == 1 || val == 0)) {
3220 if (!memcg_has_children(memcg))
3221 memcg->use_hierarchy = val;
3228 mutex_unlock(&memcg_create_mutex);
3233 static unsigned long tree_stat(struct mem_cgroup *memcg,
3234 enum mem_cgroup_stat_index idx)
3236 struct mem_cgroup *iter;
3239 /* Per-cpu values can be negative, use a signed accumulator */
3240 for_each_mem_cgroup_tree(iter, memcg)
3241 val += mem_cgroup_read_stat(iter, idx);
3243 if (val < 0) /* race ? */
3248 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3252 if (mem_cgroup_is_root(memcg)) {
3253 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3254 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3256 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3259 val = page_counter_read(&memcg->memory);
3261 val = page_counter_read(&memcg->memsw);
3263 return val << PAGE_SHIFT;
3274 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3277 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3278 struct page_counter *counter;
3280 switch (MEMFILE_TYPE(cft->private)) {
3282 counter = &memcg->memory;
3285 counter = &memcg->memsw;
3288 counter = &memcg->kmem;
3294 switch (MEMFILE_ATTR(cft->private)) {
3296 if (counter == &memcg->memory)
3297 return mem_cgroup_usage(memcg, false);
3298 if (counter == &memcg->memsw)
3299 return mem_cgroup_usage(memcg, true);
3300 return (u64)page_counter_read(counter) * PAGE_SIZE;
3302 return (u64)counter->limit * PAGE_SIZE;
3304 return (u64)counter->watermark * PAGE_SIZE;
3306 return counter->failcnt;
3307 case RES_SOFT_LIMIT:
3308 return (u64)memcg->soft_limit * PAGE_SIZE;
3314 #ifdef CONFIG_MEMCG_KMEM
3315 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3316 unsigned long nr_pages)
3321 BUG_ON(memcg->kmemcg_id >= 0);
3322 BUG_ON(memcg->kmem_acct_activated);
3323 BUG_ON(memcg->kmem_acct_active);
3326 * For simplicity, we won't allow this to be disabled. It also can't
3327 * be changed if the cgroup has children already, or if tasks had
3330 * If tasks join before we set the limit, a person looking at
3331 * kmem.usage_in_bytes will have no way to determine when it took
3332 * place, which makes the value quite meaningless.
3334 * After it first became limited, changes in the value of the limit are
3335 * of course permitted.
3337 mutex_lock(&memcg_create_mutex);
3338 if (cgroup_has_tasks(memcg->css.cgroup) ||
3339 (memcg->use_hierarchy && memcg_has_children(memcg)))
3341 mutex_unlock(&memcg_create_mutex);
3345 memcg_id = memcg_alloc_cache_id();
3352 * We couldn't have accounted to this cgroup, because it hasn't got
3353 * activated yet, so this should succeed.
3355 err = page_counter_limit(&memcg->kmem, nr_pages);
3358 static_key_slow_inc(&memcg_kmem_enabled_key);
3360 * A memory cgroup is considered kmem-active as soon as it gets
3361 * kmemcg_id. Setting the id after enabling static branching will
3362 * guarantee no one starts accounting before all call sites are
3365 memcg->kmemcg_id = memcg_id;
3366 memcg->kmem_acct_activated = true;
3367 memcg->kmem_acct_active = true;
3372 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3373 unsigned long limit)
3377 mutex_lock(&memcg_limit_mutex);
3378 if (!memcg_kmem_is_active(memcg))
3379 ret = memcg_activate_kmem(memcg, limit);
3381 ret = page_counter_limit(&memcg->kmem, limit);
3382 mutex_unlock(&memcg_limit_mutex);
3386 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3389 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3394 mutex_lock(&memcg_limit_mutex);
3396 * If the parent cgroup is not kmem-active now, it cannot be activated
3397 * after this point, because it has at least one child already.
3399 if (memcg_kmem_is_active(parent))
3400 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3401 mutex_unlock(&memcg_limit_mutex);
3405 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3406 unsigned long limit)
3410 #endif /* CONFIG_MEMCG_KMEM */
3413 * The user of this function is...
3416 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3417 char *buf, size_t nbytes, loff_t off)
3419 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3420 unsigned long nr_pages;
3423 buf = strstrip(buf);
3424 ret = page_counter_memparse(buf, "-1", &nr_pages);
3428 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3430 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3434 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3436 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3439 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3442 ret = memcg_update_kmem_limit(memcg, nr_pages);
3446 case RES_SOFT_LIMIT:
3447 memcg->soft_limit = nr_pages;
3451 return ret ?: nbytes;
3454 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3455 size_t nbytes, loff_t off)
3457 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3458 struct page_counter *counter;
3460 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3462 counter = &memcg->memory;
3465 counter = &memcg->memsw;
3468 counter = &memcg->kmem;
3474 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3476 page_counter_reset_watermark(counter);
3479 counter->failcnt = 0;
3488 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3491 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3495 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3496 struct cftype *cft, u64 val)
3498 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3500 if (val & ~MOVE_MASK)
3504 * No kind of locking is needed in here, because ->can_attach() will
3505 * check this value once in the beginning of the process, and then carry
3506 * on with stale data. This means that changes to this value will only
3507 * affect task migrations starting after the change.
3509 memcg->move_charge_at_immigrate = val;
3513 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3514 struct cftype *cft, u64 val)
3521 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3525 unsigned int lru_mask;
3528 static const struct numa_stat stats[] = {
3529 { "total", LRU_ALL },
3530 { "file", LRU_ALL_FILE },
3531 { "anon", LRU_ALL_ANON },
3532 { "unevictable", BIT(LRU_UNEVICTABLE) },
3534 const struct numa_stat *stat;
3537 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3539 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3540 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3541 seq_printf(m, "%s=%lu", stat->name, nr);
3542 for_each_node_state(nid, N_MEMORY) {
3543 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3545 seq_printf(m, " N%d=%lu", nid, nr);
3550 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3551 struct mem_cgroup *iter;
3554 for_each_mem_cgroup_tree(iter, memcg)
3555 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3556 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3557 for_each_node_state(nid, N_MEMORY) {
3559 for_each_mem_cgroup_tree(iter, memcg)
3560 nr += mem_cgroup_node_nr_lru_pages(
3561 iter, nid, stat->lru_mask);
3562 seq_printf(m, " N%d=%lu", nid, nr);
3569 #endif /* CONFIG_NUMA */
3571 static int memcg_stat_show(struct seq_file *m, void *v)
3573 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3574 unsigned long memory, memsw;
3575 struct mem_cgroup *mi;
3578 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3579 MEM_CGROUP_STAT_NSTATS);
3580 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3581 MEM_CGROUP_EVENTS_NSTATS);
3582 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3584 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3585 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3587 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3588 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3591 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3592 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3593 mem_cgroup_read_events(memcg, i));
3595 for (i = 0; i < NR_LRU_LISTS; i++)
3596 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3597 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3599 /* Hierarchical information */
3600 memory = memsw = PAGE_COUNTER_MAX;
3601 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3602 memory = min(memory, mi->memory.limit);
3603 memsw = min(memsw, mi->memsw.limit);
3605 seq_printf(m, "hierarchical_memory_limit %llu\n",
3606 (u64)memory * PAGE_SIZE);
3607 if (do_swap_account)
3608 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3609 (u64)memsw * PAGE_SIZE);
3611 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3614 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3616 for_each_mem_cgroup_tree(mi, memcg)
3617 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3618 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3621 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3622 unsigned long long val = 0;
3624 for_each_mem_cgroup_tree(mi, memcg)
3625 val += mem_cgroup_read_events(mi, i);
3626 seq_printf(m, "total_%s %llu\n",
3627 mem_cgroup_events_names[i], val);
3630 for (i = 0; i < NR_LRU_LISTS; i++) {
3631 unsigned long long val = 0;
3633 for_each_mem_cgroup_tree(mi, memcg)
3634 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3635 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3638 #ifdef CONFIG_DEBUG_VM
3641 struct mem_cgroup_per_zone *mz;
3642 struct zone_reclaim_stat *rstat;
3643 unsigned long recent_rotated[2] = {0, 0};
3644 unsigned long recent_scanned[2] = {0, 0};
3646 for_each_online_node(nid)
3647 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3648 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3649 rstat = &mz->lruvec.reclaim_stat;
3651 recent_rotated[0] += rstat->recent_rotated[0];
3652 recent_rotated[1] += rstat->recent_rotated[1];
3653 recent_scanned[0] += rstat->recent_scanned[0];
3654 recent_scanned[1] += rstat->recent_scanned[1];
3656 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3657 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3658 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3659 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3666 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3669 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3671 return mem_cgroup_swappiness(memcg);
3674 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3675 struct cftype *cft, u64 val)
3677 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3683 memcg->swappiness = val;
3685 vm_swappiness = val;
3690 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3692 struct mem_cgroup_threshold_ary *t;
3693 unsigned long usage;
3698 t = rcu_dereference(memcg->thresholds.primary);
3700 t = rcu_dereference(memcg->memsw_thresholds.primary);
3705 usage = mem_cgroup_usage(memcg, swap);
3708 * current_threshold points to threshold just below or equal to usage.
3709 * If it's not true, a threshold was crossed after last
3710 * call of __mem_cgroup_threshold().
3712 i = t->current_threshold;
3715 * Iterate backward over array of thresholds starting from
3716 * current_threshold and check if a threshold is crossed.
3717 * If none of thresholds below usage is crossed, we read
3718 * only one element of the array here.
3720 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3721 eventfd_signal(t->entries[i].eventfd, 1);
3723 /* i = current_threshold + 1 */
3727 * Iterate forward over array of thresholds starting from
3728 * current_threshold+1 and check if a threshold is crossed.
3729 * If none of thresholds above usage is crossed, we read
3730 * only one element of the array here.
3732 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3733 eventfd_signal(t->entries[i].eventfd, 1);
3735 /* Update current_threshold */
3736 t->current_threshold = i - 1;
3741 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3744 __mem_cgroup_threshold(memcg, false);
3745 if (do_swap_account)
3746 __mem_cgroup_threshold(memcg, true);
3748 memcg = parent_mem_cgroup(memcg);
3752 static int compare_thresholds(const void *a, const void *b)
3754 const struct mem_cgroup_threshold *_a = a;
3755 const struct mem_cgroup_threshold *_b = b;
3757 if (_a->threshold > _b->threshold)
3760 if (_a->threshold < _b->threshold)
3766 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3768 struct mem_cgroup_eventfd_list *ev;
3770 spin_lock(&memcg_oom_lock);
3772 list_for_each_entry(ev, &memcg->oom_notify, list)
3773 eventfd_signal(ev->eventfd, 1);
3775 spin_unlock(&memcg_oom_lock);
3779 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3781 struct mem_cgroup *iter;
3783 for_each_mem_cgroup_tree(iter, memcg)
3784 mem_cgroup_oom_notify_cb(iter);
3787 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3788 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3790 struct mem_cgroup_thresholds *thresholds;
3791 struct mem_cgroup_threshold_ary *new;
3792 unsigned long threshold;
3793 unsigned long usage;
3796 ret = page_counter_memparse(args, "-1", &threshold);
3800 mutex_lock(&memcg->thresholds_lock);
3803 thresholds = &memcg->thresholds;
3804 usage = mem_cgroup_usage(memcg, false);
3805 } else if (type == _MEMSWAP) {
3806 thresholds = &memcg->memsw_thresholds;
3807 usage = mem_cgroup_usage(memcg, true);
3811 /* Check if a threshold crossed before adding a new one */
3812 if (thresholds->primary)
3813 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3815 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3817 /* Allocate memory for new array of thresholds */
3818 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3826 /* Copy thresholds (if any) to new array */
3827 if (thresholds->primary) {
3828 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3829 sizeof(struct mem_cgroup_threshold));
3832 /* Add new threshold */
3833 new->entries[size - 1].eventfd = eventfd;
3834 new->entries[size - 1].threshold = threshold;
3836 /* Sort thresholds. Registering of new threshold isn't time-critical */
3837 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3838 compare_thresholds, NULL);
3840 /* Find current threshold */
3841 new->current_threshold = -1;
3842 for (i = 0; i < size; i++) {
3843 if (new->entries[i].threshold <= usage) {
3845 * new->current_threshold will not be used until
3846 * rcu_assign_pointer(), so it's safe to increment
3849 ++new->current_threshold;
3854 /* Free old spare buffer and save old primary buffer as spare */
3855 kfree(thresholds->spare);
3856 thresholds->spare = thresholds->primary;
3858 rcu_assign_pointer(thresholds->primary, new);
3860 /* To be sure that nobody uses thresholds */
3864 mutex_unlock(&memcg->thresholds_lock);
3869 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3870 struct eventfd_ctx *eventfd, const char *args)
3872 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3875 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3876 struct eventfd_ctx *eventfd, const char *args)
3878 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3881 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3882 struct eventfd_ctx *eventfd, enum res_type type)
3884 struct mem_cgroup_thresholds *thresholds;
3885 struct mem_cgroup_threshold_ary *new;
3886 unsigned long usage;
3889 mutex_lock(&memcg->thresholds_lock);
3892 thresholds = &memcg->thresholds;
3893 usage = mem_cgroup_usage(memcg, false);
3894 } else if (type == _MEMSWAP) {
3895 thresholds = &memcg->memsw_thresholds;
3896 usage = mem_cgroup_usage(memcg, true);
3900 if (!thresholds->primary)
3903 /* Check if a threshold crossed before removing */
3904 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3906 /* Calculate new number of threshold */
3908 for (i = 0; i < thresholds->primary->size; i++) {
3909 if (thresholds->primary->entries[i].eventfd != eventfd)
3913 new = thresholds->spare;
3915 /* Set thresholds array to NULL if we don't have thresholds */
3924 /* Copy thresholds and find current threshold */
3925 new->current_threshold = -1;
3926 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3927 if (thresholds->primary->entries[i].eventfd == eventfd)
3930 new->entries[j] = thresholds->primary->entries[i];
3931 if (new->entries[j].threshold <= usage) {
3933 * new->current_threshold will not be used
3934 * until rcu_assign_pointer(), so it's safe to increment
3937 ++new->current_threshold;
3943 /* Swap primary and spare array */
3944 thresholds->spare = thresholds->primary;
3945 /* If all events are unregistered, free the spare array */
3947 kfree(thresholds->spare);
3948 thresholds->spare = NULL;
3951 rcu_assign_pointer(thresholds->primary, new);
3953 /* To be sure that nobody uses thresholds */
3956 mutex_unlock(&memcg->thresholds_lock);
3959 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3960 struct eventfd_ctx *eventfd)
3962 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3965 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3966 struct eventfd_ctx *eventfd)
3968 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3971 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3972 struct eventfd_ctx *eventfd, const char *args)
3974 struct mem_cgroup_eventfd_list *event;
3976 event = kmalloc(sizeof(*event), GFP_KERNEL);
3980 spin_lock(&memcg_oom_lock);
3982 event->eventfd = eventfd;
3983 list_add(&event->list, &memcg->oom_notify);
3985 /* already in OOM ? */
3986 if (atomic_read(&memcg->under_oom))
3987 eventfd_signal(eventfd, 1);
3988 spin_unlock(&memcg_oom_lock);
3993 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3994 struct eventfd_ctx *eventfd)
3996 struct mem_cgroup_eventfd_list *ev, *tmp;
3998 spin_lock(&memcg_oom_lock);
4000 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4001 if (ev->eventfd == eventfd) {
4002 list_del(&ev->list);
4007 spin_unlock(&memcg_oom_lock);
4010 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4012 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4014 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4015 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4019 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4020 struct cftype *cft, u64 val)
4022 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4024 /* cannot set to root cgroup and only 0 and 1 are allowed */
4025 if (!css->parent || !((val == 0) || (val == 1)))
4028 memcg->oom_kill_disable = val;
4030 memcg_oom_recover(memcg);
4035 #ifdef CONFIG_MEMCG_KMEM
4036 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4040 ret = memcg_propagate_kmem(memcg);
4044 return mem_cgroup_sockets_init(memcg, ss);
4047 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
4049 struct cgroup_subsys_state *css;
4050 struct mem_cgroup *parent, *child;
4053 if (!memcg->kmem_acct_active)
4057 * Clear the 'active' flag before clearing memcg_caches arrays entries.
4058 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
4059 * guarantees no cache will be created for this cgroup after we are
4060 * done (see memcg_create_kmem_cache()).
4062 memcg->kmem_acct_active = false;
4064 memcg_deactivate_kmem_caches(memcg);
4066 kmemcg_id = memcg->kmemcg_id;
4067 BUG_ON(kmemcg_id < 0);
4069 parent = parent_mem_cgroup(memcg);
4071 parent = root_mem_cgroup;
4074 * Change kmemcg_id of this cgroup and all its descendants to the
4075 * parent's id, and then move all entries from this cgroup's list_lrus
4076 * to ones of the parent. After we have finished, all list_lrus
4077 * corresponding to this cgroup are guaranteed to remain empty. The
4078 * ordering is imposed by list_lru_node->lock taken by
4079 * memcg_drain_all_list_lrus().
4081 css_for_each_descendant_pre(css, &memcg->css) {
4082 child = mem_cgroup_from_css(css);
4083 BUG_ON(child->kmemcg_id != kmemcg_id);
4084 child->kmemcg_id = parent->kmemcg_id;
4085 if (!memcg->use_hierarchy)
4088 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
4090 memcg_free_cache_id(kmemcg_id);
4093 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4095 memcg_destroy_kmem_caches(memcg);
4096 mem_cgroup_sockets_destroy(memcg);
4099 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4104 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
4108 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4114 * DO NOT USE IN NEW FILES.
4116 * "cgroup.event_control" implementation.
4118 * This is way over-engineered. It tries to support fully configurable
4119 * events for each user. Such level of flexibility is completely
4120 * unnecessary especially in the light of the planned unified hierarchy.
4122 * Please deprecate this and replace with something simpler if at all
4127 * Unregister event and free resources.
4129 * Gets called from workqueue.
4131 static void memcg_event_remove(struct work_struct *work)
4133 struct mem_cgroup_event *event =
4134 container_of(work, struct mem_cgroup_event, remove);
4135 struct mem_cgroup *memcg = event->memcg;
4137 remove_wait_queue(event->wqh, &event->wait);
4139 event->unregister_event(memcg, event->eventfd);
4141 /* Notify userspace the event is going away. */
4142 eventfd_signal(event->eventfd, 1);
4144 eventfd_ctx_put(event->eventfd);
4146 css_put(&memcg->css);
4150 * Gets called on POLLHUP on eventfd when user closes it.
4152 * Called with wqh->lock held and interrupts disabled.
4154 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4155 int sync, void *key)
4157 struct mem_cgroup_event *event =
4158 container_of(wait, struct mem_cgroup_event, wait);
4159 struct mem_cgroup *memcg = event->memcg;
4160 unsigned long flags = (unsigned long)key;
4162 if (flags & POLLHUP) {
4164 * If the event has been detached at cgroup removal, we
4165 * can simply return knowing the other side will cleanup
4168 * We can't race against event freeing since the other
4169 * side will require wqh->lock via remove_wait_queue(),
4172 spin_lock(&memcg->event_list_lock);
4173 if (!list_empty(&event->list)) {
4174 list_del_init(&event->list);
4176 * We are in atomic context, but cgroup_event_remove()
4177 * may sleep, so we have to call it in workqueue.
4179 schedule_work(&event->remove);
4181 spin_unlock(&memcg->event_list_lock);
4187 static void memcg_event_ptable_queue_proc(struct file *file,
4188 wait_queue_head_t *wqh, poll_table *pt)
4190 struct mem_cgroup_event *event =
4191 container_of(pt, struct mem_cgroup_event, pt);
4194 add_wait_queue(wqh, &event->wait);
4198 * DO NOT USE IN NEW FILES.
4200 * Parse input and register new cgroup event handler.
4202 * Input must be in format '<event_fd> <control_fd> <args>'.
4203 * Interpretation of args is defined by control file implementation.
4205 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4206 char *buf, size_t nbytes, loff_t off)
4208 struct cgroup_subsys_state *css = of_css(of);
4209 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4210 struct mem_cgroup_event *event;
4211 struct cgroup_subsys_state *cfile_css;
4212 unsigned int efd, cfd;
4219 buf = strstrip(buf);
4221 efd = simple_strtoul(buf, &endp, 10);
4226 cfd = simple_strtoul(buf, &endp, 10);
4227 if ((*endp != ' ') && (*endp != '\0'))
4231 event = kzalloc(sizeof(*event), GFP_KERNEL);
4235 event->memcg = memcg;
4236 INIT_LIST_HEAD(&event->list);
4237 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4238 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4239 INIT_WORK(&event->remove, memcg_event_remove);
4247 event->eventfd = eventfd_ctx_fileget(efile.file);
4248 if (IS_ERR(event->eventfd)) {
4249 ret = PTR_ERR(event->eventfd);
4256 goto out_put_eventfd;
4259 /* the process need read permission on control file */
4260 /* AV: shouldn't we check that it's been opened for read instead? */
4261 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4266 * Determine the event callbacks and set them in @event. This used
4267 * to be done via struct cftype but cgroup core no longer knows
4268 * about these events. The following is crude but the whole thing
4269 * is for compatibility anyway.
4271 * DO NOT ADD NEW FILES.
4273 name = cfile.file->f_path.dentry->d_name.name;
4275 if (!strcmp(name, "memory.usage_in_bytes")) {
4276 event->register_event = mem_cgroup_usage_register_event;
4277 event->unregister_event = mem_cgroup_usage_unregister_event;
4278 } else if (!strcmp(name, "memory.oom_control")) {
4279 event->register_event = mem_cgroup_oom_register_event;
4280 event->unregister_event = mem_cgroup_oom_unregister_event;
4281 } else if (!strcmp(name, "memory.pressure_level")) {
4282 event->register_event = vmpressure_register_event;
4283 event->unregister_event = vmpressure_unregister_event;
4284 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4285 event->register_event = memsw_cgroup_usage_register_event;
4286 event->unregister_event = memsw_cgroup_usage_unregister_event;
4293 * Verify @cfile should belong to @css. Also, remaining events are
4294 * automatically removed on cgroup destruction but the removal is
4295 * asynchronous, so take an extra ref on @css.
4297 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4298 &memory_cgrp_subsys);
4300 if (IS_ERR(cfile_css))
4302 if (cfile_css != css) {
4307 ret = event->register_event(memcg, event->eventfd, buf);
4311 efile.file->f_op->poll(efile.file, &event->pt);
4313 spin_lock(&memcg->event_list_lock);
4314 list_add(&event->list, &memcg->event_list);
4315 spin_unlock(&memcg->event_list_lock);
4327 eventfd_ctx_put(event->eventfd);
4336 static struct cftype mem_cgroup_legacy_files[] = {
4338 .name = "usage_in_bytes",
4339 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4340 .read_u64 = mem_cgroup_read_u64,
4343 .name = "max_usage_in_bytes",
4344 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4345 .write = mem_cgroup_reset,
4346 .read_u64 = mem_cgroup_read_u64,
4349 .name = "limit_in_bytes",
4350 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4351 .write = mem_cgroup_write,
4352 .read_u64 = mem_cgroup_read_u64,
4355 .name = "soft_limit_in_bytes",
4356 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4357 .write = mem_cgroup_write,
4358 .read_u64 = mem_cgroup_read_u64,
4362 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4363 .write = mem_cgroup_reset,
4364 .read_u64 = mem_cgroup_read_u64,
4368 .seq_show = memcg_stat_show,
4371 .name = "force_empty",
4372 .write = mem_cgroup_force_empty_write,
4375 .name = "use_hierarchy",
4376 .write_u64 = mem_cgroup_hierarchy_write,
4377 .read_u64 = mem_cgroup_hierarchy_read,
4380 .name = "cgroup.event_control", /* XXX: for compat */
4381 .write = memcg_write_event_control,
4382 .flags = CFTYPE_NO_PREFIX,
4386 .name = "swappiness",
4387 .read_u64 = mem_cgroup_swappiness_read,
4388 .write_u64 = mem_cgroup_swappiness_write,
4391 .name = "move_charge_at_immigrate",
4392 .read_u64 = mem_cgroup_move_charge_read,
4393 .write_u64 = mem_cgroup_move_charge_write,
4396 .name = "oom_control",
4397 .seq_show = mem_cgroup_oom_control_read,
4398 .write_u64 = mem_cgroup_oom_control_write,
4399 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4402 .name = "pressure_level",
4406 .name = "numa_stat",
4407 .seq_show = memcg_numa_stat_show,
4410 #ifdef CONFIG_MEMCG_KMEM
4412 .name = "kmem.limit_in_bytes",
4413 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4414 .write = mem_cgroup_write,
4415 .read_u64 = mem_cgroup_read_u64,
4418 .name = "kmem.usage_in_bytes",
4419 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4420 .read_u64 = mem_cgroup_read_u64,
4423 .name = "kmem.failcnt",
4424 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4425 .write = mem_cgroup_reset,
4426 .read_u64 = mem_cgroup_read_u64,
4429 .name = "kmem.max_usage_in_bytes",
4430 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4431 .write = mem_cgroup_reset,
4432 .read_u64 = mem_cgroup_read_u64,
4434 #ifdef CONFIG_SLABINFO
4436 .name = "kmem.slabinfo",
4437 .seq_start = slab_start,
4438 .seq_next = slab_next,
4439 .seq_stop = slab_stop,
4440 .seq_show = memcg_slab_show,
4444 { }, /* terminate */
4447 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4449 struct mem_cgroup_per_node *pn;
4450 struct mem_cgroup_per_zone *mz;
4451 int zone, tmp = node;
4453 * This routine is called against possible nodes.
4454 * But it's BUG to call kmalloc() against offline node.
4456 * TODO: this routine can waste much memory for nodes which will
4457 * never be onlined. It's better to use memory hotplug callback
4460 if (!node_state(node, N_NORMAL_MEMORY))
4462 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4466 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4467 mz = &pn->zoneinfo[zone];
4468 lruvec_init(&mz->lruvec);
4469 mz->usage_in_excess = 0;
4470 mz->on_tree = false;
4473 memcg->nodeinfo[node] = pn;
4477 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4479 kfree(memcg->nodeinfo[node]);
4482 static struct mem_cgroup *mem_cgroup_alloc(void)
4484 struct mem_cgroup *memcg;
4487 size = sizeof(struct mem_cgroup);
4488 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4490 memcg = kzalloc(size, GFP_KERNEL);
4494 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4497 spin_lock_init(&memcg->pcp_counter_lock);
4506 * At destroying mem_cgroup, references from swap_cgroup can remain.
4507 * (scanning all at force_empty is too costly...)
4509 * Instead of clearing all references at force_empty, we remember
4510 * the number of reference from swap_cgroup and free mem_cgroup when
4511 * it goes down to 0.
4513 * Removal of cgroup itself succeeds regardless of refs from swap.
4516 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4520 mem_cgroup_remove_from_trees(memcg);
4523 free_mem_cgroup_per_zone_info(memcg, node);
4525 free_percpu(memcg->stat);
4527 disarm_static_keys(memcg);
4532 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4534 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4536 if (!memcg->memory.parent)
4538 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4540 EXPORT_SYMBOL(parent_mem_cgroup);
4542 static struct cgroup_subsys_state * __ref
4543 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4545 struct mem_cgroup *memcg;
4546 long error = -ENOMEM;
4549 memcg = mem_cgroup_alloc();
4551 return ERR_PTR(error);
4554 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4558 if (parent_css == NULL) {
4559 root_mem_cgroup = memcg;
4560 page_counter_init(&memcg->memory, NULL);
4561 memcg->high = PAGE_COUNTER_MAX;
4562 memcg->soft_limit = PAGE_COUNTER_MAX;
4563 page_counter_init(&memcg->memsw, NULL);
4564 page_counter_init(&memcg->kmem, NULL);
4567 memcg->last_scanned_node = MAX_NUMNODES;
4568 INIT_LIST_HEAD(&memcg->oom_notify);
4569 memcg->move_charge_at_immigrate = 0;
4570 mutex_init(&memcg->thresholds_lock);
4571 spin_lock_init(&memcg->move_lock);
4572 vmpressure_init(&memcg->vmpressure);
4573 INIT_LIST_HEAD(&memcg->event_list);
4574 spin_lock_init(&memcg->event_list_lock);
4575 #ifdef CONFIG_MEMCG_KMEM
4576 memcg->kmemcg_id = -1;
4582 __mem_cgroup_free(memcg);
4583 return ERR_PTR(error);
4587 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4589 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4590 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4593 if (css->id > MEM_CGROUP_ID_MAX)
4599 mutex_lock(&memcg_create_mutex);
4601 memcg->use_hierarchy = parent->use_hierarchy;
4602 memcg->oom_kill_disable = parent->oom_kill_disable;
4603 memcg->swappiness = mem_cgroup_swappiness(parent);
4605 if (parent->use_hierarchy) {
4606 page_counter_init(&memcg->memory, &parent->memory);
4607 memcg->high = PAGE_COUNTER_MAX;
4608 memcg->soft_limit = PAGE_COUNTER_MAX;
4609 page_counter_init(&memcg->memsw, &parent->memsw);
4610 page_counter_init(&memcg->kmem, &parent->kmem);
4613 * No need to take a reference to the parent because cgroup
4614 * core guarantees its existence.
4617 page_counter_init(&memcg->memory, NULL);
4618 memcg->high = PAGE_COUNTER_MAX;
4619 memcg->soft_limit = PAGE_COUNTER_MAX;
4620 page_counter_init(&memcg->memsw, NULL);
4621 page_counter_init(&memcg->kmem, NULL);
4623 * Deeper hierachy with use_hierarchy == false doesn't make
4624 * much sense so let cgroup subsystem know about this
4625 * unfortunate state in our controller.
4627 if (parent != root_mem_cgroup)
4628 memory_cgrp_subsys.broken_hierarchy = true;
4630 mutex_unlock(&memcg_create_mutex);
4632 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4637 * Make sure the memcg is initialized: mem_cgroup_iter()
4638 * orders reading memcg->initialized against its callers
4639 * reading the memcg members.
4641 smp_store_release(&memcg->initialized, 1);
4646 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4648 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4649 struct mem_cgroup_event *event, *tmp;
4652 * Unregister events and notify userspace.
4653 * Notify userspace about cgroup removing only after rmdir of cgroup
4654 * directory to avoid race between userspace and kernelspace.
4656 spin_lock(&memcg->event_list_lock);
4657 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4658 list_del_init(&event->list);
4659 schedule_work(&event->remove);
4661 spin_unlock(&memcg->event_list_lock);
4663 vmpressure_cleanup(&memcg->vmpressure);
4665 memcg_deactivate_kmem(memcg);
4668 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4670 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4672 memcg_destroy_kmem(memcg);
4673 __mem_cgroup_free(memcg);
4677 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4678 * @css: the target css
4680 * Reset the states of the mem_cgroup associated with @css. This is
4681 * invoked when the userland requests disabling on the default hierarchy
4682 * but the memcg is pinned through dependency. The memcg should stop
4683 * applying policies and should revert to the vanilla state as it may be
4684 * made visible again.
4686 * The current implementation only resets the essential configurations.
4687 * This needs to be expanded to cover all the visible parts.
4689 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4691 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4693 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4694 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4695 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4697 memcg->high = PAGE_COUNTER_MAX;
4698 memcg->soft_limit = PAGE_COUNTER_MAX;
4702 /* Handlers for move charge at task migration. */
4703 static int mem_cgroup_do_precharge(unsigned long count)
4707 /* Try a single bulk charge without reclaim first */
4708 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4710 mc.precharge += count;
4713 if (ret == -EINTR) {
4714 cancel_charge(root_mem_cgroup, count);
4718 /* Try charges one by one with reclaim */
4720 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4722 * In case of failure, any residual charges against
4723 * mc.to will be dropped by mem_cgroup_clear_mc()
4724 * later on. However, cancel any charges that are
4725 * bypassed to root right away or they'll be lost.
4728 cancel_charge(root_mem_cgroup, 1);
4738 * get_mctgt_type - get target type of moving charge
4739 * @vma: the vma the pte to be checked belongs
4740 * @addr: the address corresponding to the pte to be checked
4741 * @ptent: the pte to be checked
4742 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4745 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4746 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4747 * move charge. if @target is not NULL, the page is stored in target->page
4748 * with extra refcnt got(Callers should handle it).
4749 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4750 * target for charge migration. if @target is not NULL, the entry is stored
4753 * Called with pte lock held.
4760 enum mc_target_type {
4766 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4767 unsigned long addr, pte_t ptent)
4769 struct page *page = vm_normal_page(vma, addr, ptent);
4771 if (!page || !page_mapped(page))
4773 if (PageAnon(page)) {
4774 if (!(mc.flags & MOVE_ANON))
4777 if (!(mc.flags & MOVE_FILE))
4780 if (!get_page_unless_zero(page))
4787 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4788 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4790 struct page *page = NULL;
4791 swp_entry_t ent = pte_to_swp_entry(ptent);
4793 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4796 * Because lookup_swap_cache() updates some statistics counter,
4797 * we call find_get_page() with swapper_space directly.
4799 page = find_get_page(swap_address_space(ent), ent.val);
4800 if (do_swap_account)
4801 entry->val = ent.val;
4806 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4807 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4813 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4814 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4816 struct page *page = NULL;
4817 struct address_space *mapping;
4820 if (!vma->vm_file) /* anonymous vma */
4822 if (!(mc.flags & MOVE_FILE))
4825 mapping = vma->vm_file->f_mapping;
4826 pgoff = linear_page_index(vma, addr);
4828 /* page is moved even if it's not RSS of this task(page-faulted). */
4830 /* shmem/tmpfs may report page out on swap: account for that too. */
4831 if (shmem_mapping(mapping)) {
4832 page = find_get_entry(mapping, pgoff);
4833 if (radix_tree_exceptional_entry(page)) {
4834 swp_entry_t swp = radix_to_swp_entry(page);
4835 if (do_swap_account)
4837 page = find_get_page(swap_address_space(swp), swp.val);
4840 page = find_get_page(mapping, pgoff);
4842 page = find_get_page(mapping, pgoff);
4847 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4848 unsigned long addr, pte_t ptent, union mc_target *target)
4850 struct page *page = NULL;
4851 enum mc_target_type ret = MC_TARGET_NONE;
4852 swp_entry_t ent = { .val = 0 };
4854 if (pte_present(ptent))
4855 page = mc_handle_present_pte(vma, addr, ptent);
4856 else if (is_swap_pte(ptent))
4857 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4858 else if (pte_none(ptent))
4859 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4861 if (!page && !ent.val)
4865 * Do only loose check w/o serialization.
4866 * mem_cgroup_move_account() checks the page is valid or
4867 * not under LRU exclusion.
4869 if (page->mem_cgroup == mc.from) {
4870 ret = MC_TARGET_PAGE;
4872 target->page = page;
4874 if (!ret || !target)
4877 /* There is a swap entry and a page doesn't exist or isn't charged */
4878 if (ent.val && !ret &&
4879 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4880 ret = MC_TARGET_SWAP;
4887 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4889 * We don't consider swapping or file mapped pages because THP does not
4890 * support them for now.
4891 * Caller should make sure that pmd_trans_huge(pmd) is true.
4893 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4894 unsigned long addr, pmd_t pmd, union mc_target *target)
4896 struct page *page = NULL;
4897 enum mc_target_type ret = MC_TARGET_NONE;
4899 page = pmd_page(pmd);
4900 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4901 if (!(mc.flags & MOVE_ANON))
4903 if (page->mem_cgroup == mc.from) {
4904 ret = MC_TARGET_PAGE;
4907 target->page = page;
4913 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4914 unsigned long addr, pmd_t pmd, union mc_target *target)
4916 return MC_TARGET_NONE;
4920 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4921 unsigned long addr, unsigned long end,
4922 struct mm_walk *walk)
4924 struct vm_area_struct *vma = walk->vma;
4928 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4929 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4930 mc.precharge += HPAGE_PMD_NR;
4935 if (pmd_trans_unstable(pmd))
4937 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4938 for (; addr != end; pte++, addr += PAGE_SIZE)
4939 if (get_mctgt_type(vma, addr, *pte, NULL))
4940 mc.precharge++; /* increment precharge temporarily */
4941 pte_unmap_unlock(pte - 1, ptl);
4947 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4949 unsigned long precharge;
4951 struct mm_walk mem_cgroup_count_precharge_walk = {
4952 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4955 down_read(&mm->mmap_sem);
4956 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4957 up_read(&mm->mmap_sem);
4959 precharge = mc.precharge;
4965 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4967 unsigned long precharge = mem_cgroup_count_precharge(mm);
4969 VM_BUG_ON(mc.moving_task);
4970 mc.moving_task = current;
4971 return mem_cgroup_do_precharge(precharge);
4974 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4975 static void __mem_cgroup_clear_mc(void)
4977 struct mem_cgroup *from = mc.from;
4978 struct mem_cgroup *to = mc.to;
4980 /* we must uncharge all the leftover precharges from mc.to */
4982 cancel_charge(mc.to, mc.precharge);
4986 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4987 * we must uncharge here.
4989 if (mc.moved_charge) {
4990 cancel_charge(mc.from, mc.moved_charge);
4991 mc.moved_charge = 0;
4993 /* we must fixup refcnts and charges */
4994 if (mc.moved_swap) {
4995 /* uncharge swap account from the old cgroup */
4996 if (!mem_cgroup_is_root(mc.from))
4997 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5000 * we charged both to->memory and to->memsw, so we
5001 * should uncharge to->memory.
5003 if (!mem_cgroup_is_root(mc.to))
5004 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5006 css_put_many(&mc.from->css, mc.moved_swap);
5008 /* we've already done css_get(mc.to) */
5011 memcg_oom_recover(from);
5012 memcg_oom_recover(to);
5013 wake_up_all(&mc.waitq);
5016 static void mem_cgroup_clear_mc(void)
5019 * we must clear moving_task before waking up waiters at the end of
5022 mc.moving_task = NULL;
5023 __mem_cgroup_clear_mc();
5024 spin_lock(&mc.lock);
5027 spin_unlock(&mc.lock);
5030 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5031 struct cgroup_taskset *tset)
5033 struct task_struct *p = cgroup_taskset_first(tset);
5035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5036 unsigned long move_flags;
5039 * We are now commited to this value whatever it is. Changes in this
5040 * tunable will only affect upcoming migrations, not the current one.
5041 * So we need to save it, and keep it going.
5043 move_flags = ACCESS_ONCE(memcg->move_charge_at_immigrate);
5045 struct mm_struct *mm;
5046 struct mem_cgroup *from = mem_cgroup_from_task(p);
5048 VM_BUG_ON(from == memcg);
5050 mm = get_task_mm(p);
5053 /* We move charges only when we move a owner of the mm */
5054 if (mm->owner == p) {
5057 VM_BUG_ON(mc.precharge);
5058 VM_BUG_ON(mc.moved_charge);
5059 VM_BUG_ON(mc.moved_swap);
5061 spin_lock(&mc.lock);
5064 mc.flags = move_flags;
5065 spin_unlock(&mc.lock);
5066 /* We set mc.moving_task later */
5068 ret = mem_cgroup_precharge_mc(mm);
5070 mem_cgroup_clear_mc();
5077 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5078 struct cgroup_taskset *tset)
5081 mem_cgroup_clear_mc();
5084 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5085 unsigned long addr, unsigned long end,
5086 struct mm_walk *walk)
5089 struct vm_area_struct *vma = walk->vma;
5092 enum mc_target_type target_type;
5093 union mc_target target;
5097 * We don't take compound_lock() here but no race with splitting thp
5099 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5100 * under splitting, which means there's no concurrent thp split,
5101 * - if another thread runs into split_huge_page() just after we
5102 * entered this if-block, the thread must wait for page table lock
5103 * to be unlocked in __split_huge_page_splitting(), where the main
5104 * part of thp split is not executed yet.
5106 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5107 if (mc.precharge < HPAGE_PMD_NR) {
5111 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5112 if (target_type == MC_TARGET_PAGE) {
5114 if (!isolate_lru_page(page)) {
5115 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5117 mc.precharge -= HPAGE_PMD_NR;
5118 mc.moved_charge += HPAGE_PMD_NR;
5120 putback_lru_page(page);
5128 if (pmd_trans_unstable(pmd))
5131 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5132 for (; addr != end; addr += PAGE_SIZE) {
5133 pte_t ptent = *(pte++);
5139 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5140 case MC_TARGET_PAGE:
5142 if (isolate_lru_page(page))
5144 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5146 /* we uncharge from mc.from later. */
5149 putback_lru_page(page);
5150 put: /* get_mctgt_type() gets the page */
5153 case MC_TARGET_SWAP:
5155 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5157 /* we fixup refcnts and charges later. */
5165 pte_unmap_unlock(pte - 1, ptl);
5170 * We have consumed all precharges we got in can_attach().
5171 * We try charge one by one, but don't do any additional
5172 * charges to mc.to if we have failed in charge once in attach()
5175 ret = mem_cgroup_do_precharge(1);
5183 static void mem_cgroup_move_charge(struct mm_struct *mm)
5185 struct mm_walk mem_cgroup_move_charge_walk = {
5186 .pmd_entry = mem_cgroup_move_charge_pte_range,
5190 lru_add_drain_all();
5192 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5193 * move_lock while we're moving its pages to another memcg.
5194 * Then wait for already started RCU-only updates to finish.
5196 atomic_inc(&mc.from->moving_account);
5199 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5201 * Someone who are holding the mmap_sem might be waiting in
5202 * waitq. So we cancel all extra charges, wake up all waiters,
5203 * and retry. Because we cancel precharges, we might not be able
5204 * to move enough charges, but moving charge is a best-effort
5205 * feature anyway, so it wouldn't be a big problem.
5207 __mem_cgroup_clear_mc();
5212 * When we have consumed all precharges and failed in doing
5213 * additional charge, the page walk just aborts.
5215 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5216 up_read(&mm->mmap_sem);
5217 atomic_dec(&mc.from->moving_account);
5220 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5221 struct cgroup_taskset *tset)
5223 struct task_struct *p = cgroup_taskset_first(tset);
5224 struct mm_struct *mm = get_task_mm(p);
5228 mem_cgroup_move_charge(mm);
5232 mem_cgroup_clear_mc();
5234 #else /* !CONFIG_MMU */
5235 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5236 struct cgroup_taskset *tset)
5240 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5241 struct cgroup_taskset *tset)
5244 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5245 struct cgroup_taskset *tset)
5251 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5252 * to verify whether we're attached to the default hierarchy on each mount
5255 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5258 * use_hierarchy is forced on the default hierarchy. cgroup core
5259 * guarantees that @root doesn't have any children, so turning it
5260 * on for the root memcg is enough.
5262 if (cgroup_on_dfl(root_css->cgroup))
5263 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5266 static u64 memory_current_read(struct cgroup_subsys_state *css,
5269 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5272 static int memory_low_show(struct seq_file *m, void *v)
5274 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5275 unsigned long low = ACCESS_ONCE(memcg->low);
5277 if (low == PAGE_COUNTER_MAX)
5278 seq_puts(m, "infinity\n");
5280 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5285 static ssize_t memory_low_write(struct kernfs_open_file *of,
5286 char *buf, size_t nbytes, loff_t off)
5288 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5292 buf = strstrip(buf);
5293 err = page_counter_memparse(buf, "infinity", &low);
5302 static int memory_high_show(struct seq_file *m, void *v)
5304 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5305 unsigned long high = ACCESS_ONCE(memcg->high);
5307 if (high == PAGE_COUNTER_MAX)
5308 seq_puts(m, "infinity\n");
5310 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5315 static ssize_t memory_high_write(struct kernfs_open_file *of,
5316 char *buf, size_t nbytes, loff_t off)
5318 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5322 buf = strstrip(buf);
5323 err = page_counter_memparse(buf, "infinity", &high);
5332 static int memory_max_show(struct seq_file *m, void *v)
5334 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5335 unsigned long max = ACCESS_ONCE(memcg->memory.limit);
5337 if (max == PAGE_COUNTER_MAX)
5338 seq_puts(m, "infinity\n");
5340 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5345 static ssize_t memory_max_write(struct kernfs_open_file *of,
5346 char *buf, size_t nbytes, loff_t off)
5348 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5352 buf = strstrip(buf);
5353 err = page_counter_memparse(buf, "infinity", &max);
5357 err = mem_cgroup_resize_limit(memcg, max);
5364 static int memory_events_show(struct seq_file *m, void *v)
5366 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5368 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5369 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5370 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5371 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5376 static struct cftype memory_files[] = {
5379 .read_u64 = memory_current_read,
5383 .flags = CFTYPE_NOT_ON_ROOT,
5384 .seq_show = memory_low_show,
5385 .write = memory_low_write,
5389 .flags = CFTYPE_NOT_ON_ROOT,
5390 .seq_show = memory_high_show,
5391 .write = memory_high_write,
5395 .flags = CFTYPE_NOT_ON_ROOT,
5396 .seq_show = memory_max_show,
5397 .write = memory_max_write,
5401 .flags = CFTYPE_NOT_ON_ROOT,
5402 .seq_show = memory_events_show,
5407 struct cgroup_subsys memory_cgrp_subsys = {
5408 .css_alloc = mem_cgroup_css_alloc,
5409 .css_online = mem_cgroup_css_online,
5410 .css_offline = mem_cgroup_css_offline,
5411 .css_free = mem_cgroup_css_free,
5412 .css_reset = mem_cgroup_css_reset,
5413 .can_attach = mem_cgroup_can_attach,
5414 .cancel_attach = mem_cgroup_cancel_attach,
5415 .attach = mem_cgroup_move_task,
5416 .bind = mem_cgroup_bind,
5417 .dfl_cftypes = memory_files,
5418 .legacy_cftypes = mem_cgroup_legacy_files,
5423 * mem_cgroup_events - count memory events against a cgroup
5424 * @memcg: the memory cgroup
5425 * @idx: the event index
5426 * @nr: the number of events to account for
5428 void mem_cgroup_events(struct mem_cgroup *memcg,
5429 enum mem_cgroup_events_index idx,
5432 this_cpu_add(memcg->stat->events[idx], nr);
5436 * mem_cgroup_low - check if memory consumption is below the normal range
5437 * @root: the highest ancestor to consider
5438 * @memcg: the memory cgroup to check
5440 * Returns %true if memory consumption of @memcg, and that of all
5441 * configurable ancestors up to @root, is below the normal range.
5443 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5445 if (mem_cgroup_disabled())
5449 * The toplevel group doesn't have a configurable range, so
5450 * it's never low when looked at directly, and it is not
5451 * considered an ancestor when assessing the hierarchy.
5454 if (memcg == root_mem_cgroup)
5457 if (page_counter_read(&memcg->memory) > memcg->low)
5460 while (memcg != root) {
5461 memcg = parent_mem_cgroup(memcg);
5463 if (memcg == root_mem_cgroup)
5466 if (page_counter_read(&memcg->memory) > memcg->low)
5473 * mem_cgroup_try_charge - try charging a page
5474 * @page: page to charge
5475 * @mm: mm context of the victim
5476 * @gfp_mask: reclaim mode
5477 * @memcgp: charged memcg return
5479 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5480 * pages according to @gfp_mask if necessary.
5482 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5483 * Otherwise, an error code is returned.
5485 * After page->mapping has been set up, the caller must finalize the
5486 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5487 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5489 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5490 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5492 struct mem_cgroup *memcg = NULL;
5493 unsigned int nr_pages = 1;
5496 if (mem_cgroup_disabled())
5499 if (PageSwapCache(page)) {
5501 * Every swap fault against a single page tries to charge the
5502 * page, bail as early as possible. shmem_unuse() encounters
5503 * already charged pages, too. The USED bit is protected by
5504 * the page lock, which serializes swap cache removal, which
5505 * in turn serializes uncharging.
5507 if (page->mem_cgroup)
5511 if (PageTransHuge(page)) {
5512 nr_pages <<= compound_order(page);
5513 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5516 if (do_swap_account && PageSwapCache(page))
5517 memcg = try_get_mem_cgroup_from_page(page);
5519 memcg = get_mem_cgroup_from_mm(mm);
5521 ret = try_charge(memcg, gfp_mask, nr_pages);
5523 css_put(&memcg->css);
5525 if (ret == -EINTR) {
5526 memcg = root_mem_cgroup;
5535 * mem_cgroup_commit_charge - commit a page charge
5536 * @page: page to charge
5537 * @memcg: memcg to charge the page to
5538 * @lrucare: page might be on LRU already
5540 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5541 * after page->mapping has been set up. This must happen atomically
5542 * as part of the page instantiation, i.e. under the page table lock
5543 * for anonymous pages, under the page lock for page and swap cache.
5545 * In addition, the page must not be on the LRU during the commit, to
5546 * prevent racing with task migration. If it might be, use @lrucare.
5548 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5550 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5553 unsigned int nr_pages = 1;
5555 VM_BUG_ON_PAGE(!page->mapping, page);
5556 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5558 if (mem_cgroup_disabled())
5561 * Swap faults will attempt to charge the same page multiple
5562 * times. But reuse_swap_page() might have removed the page
5563 * from swapcache already, so we can't check PageSwapCache().
5568 commit_charge(page, memcg, lrucare);
5570 if (PageTransHuge(page)) {
5571 nr_pages <<= compound_order(page);
5572 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5575 local_irq_disable();
5576 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5577 memcg_check_events(memcg, page);
5580 if (do_swap_account && PageSwapCache(page)) {
5581 swp_entry_t entry = { .val = page_private(page) };
5583 * The swap entry might not get freed for a long time,
5584 * let's not wait for it. The page already received a
5585 * memory+swap charge, drop the swap entry duplicate.
5587 mem_cgroup_uncharge_swap(entry);
5592 * mem_cgroup_cancel_charge - cancel a page charge
5593 * @page: page to charge
5594 * @memcg: memcg to charge the page to
5596 * Cancel a charge transaction started by mem_cgroup_try_charge().
5598 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5600 unsigned int nr_pages = 1;
5602 if (mem_cgroup_disabled())
5605 * Swap faults will attempt to charge the same page multiple
5606 * times. But reuse_swap_page() might have removed the page
5607 * from swapcache already, so we can't check PageSwapCache().
5612 if (PageTransHuge(page)) {
5613 nr_pages <<= compound_order(page);
5614 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5617 cancel_charge(memcg, nr_pages);
5620 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5621 unsigned long nr_anon, unsigned long nr_file,
5622 unsigned long nr_huge, struct page *dummy_page)
5624 unsigned long nr_pages = nr_anon + nr_file;
5625 unsigned long flags;
5627 if (!mem_cgroup_is_root(memcg)) {
5628 page_counter_uncharge(&memcg->memory, nr_pages);
5629 if (do_swap_account)
5630 page_counter_uncharge(&memcg->memsw, nr_pages);
5631 memcg_oom_recover(memcg);
5634 local_irq_save(flags);
5635 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5636 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5637 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5638 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5639 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5640 memcg_check_events(memcg, dummy_page);
5641 local_irq_restore(flags);
5643 if (!mem_cgroup_is_root(memcg))
5644 css_put_many(&memcg->css, nr_pages);
5647 static void uncharge_list(struct list_head *page_list)
5649 struct mem_cgroup *memcg = NULL;
5650 unsigned long nr_anon = 0;
5651 unsigned long nr_file = 0;
5652 unsigned long nr_huge = 0;
5653 unsigned long pgpgout = 0;
5654 struct list_head *next;
5657 next = page_list->next;
5659 unsigned int nr_pages = 1;
5661 page = list_entry(next, struct page, lru);
5662 next = page->lru.next;
5664 VM_BUG_ON_PAGE(PageLRU(page), page);
5665 VM_BUG_ON_PAGE(page_count(page), page);
5667 if (!page->mem_cgroup)
5671 * Nobody should be changing or seriously looking at
5672 * page->mem_cgroup at this point, we have fully
5673 * exclusive access to the page.
5676 if (memcg != page->mem_cgroup) {
5678 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5680 pgpgout = nr_anon = nr_file = nr_huge = 0;
5682 memcg = page->mem_cgroup;
5685 if (PageTransHuge(page)) {
5686 nr_pages <<= compound_order(page);
5687 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5688 nr_huge += nr_pages;
5692 nr_anon += nr_pages;
5694 nr_file += nr_pages;
5696 page->mem_cgroup = NULL;
5699 } while (next != page_list);
5702 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5707 * mem_cgroup_uncharge - uncharge a page
5708 * @page: page to uncharge
5710 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5711 * mem_cgroup_commit_charge().
5713 void mem_cgroup_uncharge(struct page *page)
5715 if (mem_cgroup_disabled())
5718 /* Don't touch page->lru of any random page, pre-check: */
5719 if (!page->mem_cgroup)
5722 INIT_LIST_HEAD(&page->lru);
5723 uncharge_list(&page->lru);
5727 * mem_cgroup_uncharge_list - uncharge a list of page
5728 * @page_list: list of pages to uncharge
5730 * Uncharge a list of pages previously charged with
5731 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5733 void mem_cgroup_uncharge_list(struct list_head *page_list)
5735 if (mem_cgroup_disabled())
5738 if (!list_empty(page_list))
5739 uncharge_list(page_list);
5743 * mem_cgroup_migrate - migrate a charge to another page
5744 * @oldpage: currently charged page
5745 * @newpage: page to transfer the charge to
5746 * @lrucare: either or both pages might be on the LRU already
5748 * Migrate the charge from @oldpage to @newpage.
5750 * Both pages must be locked, @newpage->mapping must be set up.
5752 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5755 struct mem_cgroup *memcg;
5758 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5759 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5760 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5761 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5762 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5763 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5766 if (mem_cgroup_disabled())
5769 /* Page cache replacement: new page already charged? */
5770 if (newpage->mem_cgroup)
5774 * Swapcache readahead pages can get migrated before being
5775 * charged, and migration from compaction can happen to an
5776 * uncharged page when the PFN walker finds a page that
5777 * reclaim just put back on the LRU but has not released yet.
5779 memcg = oldpage->mem_cgroup;
5784 lock_page_lru(oldpage, &isolated);
5786 oldpage->mem_cgroup = NULL;
5789 unlock_page_lru(oldpage, isolated);
5791 commit_charge(newpage, memcg, lrucare);
5795 * subsys_initcall() for memory controller.
5797 * Some parts like hotcpu_notifier() have to be initialized from this context
5798 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5799 * everything that doesn't depend on a specific mem_cgroup structure should
5800 * be initialized from here.
5802 static int __init mem_cgroup_init(void)
5806 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5808 for_each_possible_cpu(cpu)
5809 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5812 for_each_node(node) {
5813 struct mem_cgroup_tree_per_node *rtpn;
5816 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5817 node_online(node) ? node : NUMA_NO_NODE);
5819 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5820 struct mem_cgroup_tree_per_zone *rtpz;
5822 rtpz = &rtpn->rb_tree_per_zone[zone];
5823 rtpz->rb_root = RB_ROOT;
5824 spin_lock_init(&rtpz->lock);
5826 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5831 subsys_initcall(mem_cgroup_init);
5833 #ifdef CONFIG_MEMCG_SWAP
5835 * mem_cgroup_swapout - transfer a memsw charge to swap
5836 * @page: page whose memsw charge to transfer
5837 * @entry: swap entry to move the charge to
5839 * Transfer the memsw charge of @page to @entry.
5841 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5843 struct mem_cgroup *memcg;
5844 unsigned short oldid;
5846 VM_BUG_ON_PAGE(PageLRU(page), page);
5847 VM_BUG_ON_PAGE(page_count(page), page);
5849 if (!do_swap_account)
5852 memcg = page->mem_cgroup;
5854 /* Readahead page, never charged */
5858 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5859 VM_BUG_ON_PAGE(oldid, page);
5860 mem_cgroup_swap_statistics(memcg, true);
5862 page->mem_cgroup = NULL;
5864 if (!mem_cgroup_is_root(memcg))
5865 page_counter_uncharge(&memcg->memory, 1);
5867 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5868 VM_BUG_ON(!irqs_disabled());
5870 mem_cgroup_charge_statistics(memcg, page, -1);
5871 memcg_check_events(memcg, page);
5875 * mem_cgroup_uncharge_swap - uncharge a swap entry
5876 * @entry: swap entry to uncharge
5878 * Drop the memsw charge associated with @entry.
5880 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5882 struct mem_cgroup *memcg;
5885 if (!do_swap_account)
5888 id = swap_cgroup_record(entry, 0);
5890 memcg = mem_cgroup_lookup(id);
5892 if (!mem_cgroup_is_root(memcg))
5893 page_counter_uncharge(&memcg->memsw, 1);
5894 mem_cgroup_swap_statistics(memcg, false);
5895 css_put(&memcg->css);
5900 /* for remember boot option*/
5901 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5902 static int really_do_swap_account __initdata = 1;
5904 static int really_do_swap_account __initdata;
5907 static int __init enable_swap_account(char *s)
5909 if (!strcmp(s, "1"))
5910 really_do_swap_account = 1;
5911 else if (!strcmp(s, "0"))
5912 really_do_swap_account = 0;
5915 __setup("swapaccount=", enable_swap_account);
5917 static struct cftype memsw_cgroup_files[] = {
5919 .name = "memsw.usage_in_bytes",
5920 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5921 .read_u64 = mem_cgroup_read_u64,
5924 .name = "memsw.max_usage_in_bytes",
5925 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5926 .write = mem_cgroup_reset,
5927 .read_u64 = mem_cgroup_read_u64,
5930 .name = "memsw.limit_in_bytes",
5931 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5932 .write = mem_cgroup_write,
5933 .read_u64 = mem_cgroup_read_u64,
5936 .name = "memsw.failcnt",
5937 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5938 .write = mem_cgroup_reset,
5939 .read_u64 = mem_cgroup_read_u64,
5941 { }, /* terminate */
5944 static int __init mem_cgroup_swap_init(void)
5946 if (!mem_cgroup_disabled() && really_do_swap_account) {
5947 do_swap_account = 1;
5948 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5949 memsw_cgroup_files));
5953 subsys_initcall(mem_cgroup_swap_init);
5955 #endif /* CONFIG_MEMCG_SWAP */